I would like to introduce, through this article, a concept I have been working on for several months: semantic resistance.

The idea starts from a simple but uncomfortable observation.

You - as a company - may have spent 20 years building your differentiation. A refined purchasing process. A due diligence methodology. A proprietary design vocabulary. A candidate scoring framework. A sales playbook. An expert assessment protocol.

These intangible assets are what create your market value.

Bad news: an LLM does not possess your know-how. But it already possesses, or soon will possess, the common language in which you describe it. And it is through this common language that your differentiation gradually becomes generable, imitable, and then commoditized.

So everything you describe in standard English, standard French … already belongs to it. Or soon will, in three months, in one or two model versions.

The boundary between “proprietary know-how” and “standard generative content” is moving faster than most organizations realize.

S3m@nt1c r3s1st@nc3 proposes two defensive moves.

1. Make the language of your differentiation harder to absorb.

This does not mean becoming obscure.

It means separating two languages:

• the public language you use to be found, understood, and cited by AI systems,

• and the internal language through which your organization actually thinks, decides, scores, prioritizes, diagnoses, and acts.

GEO (Generative Engine Optimization) pushes companies to expose more and more of the first language.

The danger is when they start exposing the second one.

That is not visibility. That is self-commoditization.

When you give models a clean verbal blueprint of your deepest know-how, you are not just optimizing for AI search.

You are making your differentiation easier to absorb, reproduce, and erase.

Stop describing what you know how to do with everyone else’s words.

Forge a proprietary vocabulary. Give internal names to your methods, your steps, your categories. Change your language faster than LLM builders can assimilate it.

This is the defense of the weak against the strong.

Radio Londres (the french resistance from the UK) spoke in metaphors: “Les sanglots longs des violons de l’automne…” Incomprehensible to the occupier, perfectly clear to those who needed to understand.

The principle is exactly the same.

Universal clarity is commoditizable. A shared code among insiders is not.

Subscribe now

2. Capitalize on usage: cumulative value AND creative value.

Every time one of your key intellectual processes is run, two things happen at the same time.

You accumulate: friction data, field feedback, proprietary signals that only you see because only you are in a position to see them.

And you create: new symbols, new terms, new distinctions to name what you discover as you move forward.

Usage produces both.

This is what resists absorption in practice: not only because you see what others do not see, but because you are constantly inventing the vocabulary to describe it — faster than a model can learn it.

The same reasoning applies at the scale of a nation.

A country that only speaks the language of dominant models — the English of Californian training data — has already lost cognitive sovereignty over its own economy.

Its legal concepts, administrative categories, regulatory frameworks, and strategic thinking pass through models it did not train, does not control, and that were optimized on reference systems other than its own.

I admire Mistral’s effort, but AI sovereignty cannot be reduced to producing a European OpenAI.

AI sovereignty is not only a matter of infrastructure or chips. It is first and foremost a linguistic and conceptual issue.

Which corpora? Which language? Which business concepts specific to a given industrial fabric? Which models trained on what?

Deprive the Large Language Model of its raw material and it becomes nothing more than a naked king: a Large Commodity Model.

For any country that wants to matter in tomorrow’s economy, semantic resistance is the same battle fought at two scales: the company and the nation.

We do not protect ourselves from AI, or from the powers that dominate it — the US and China — by ignoring it.

We protect ourselves by making it impossible for AI to absorb what makes us singular.

Share

#SemanticResistance #LLM #AISovereignty

Sellers still think information is power. They're two years behind.

In my business and in my personal life, I am still receiving proposals that are grossly unbalanced (favoring the other party) and I don’t understand why.

People who I work with know I’m all in with AI.

They know that upon reception I will upload the proposal in my favorite AIs (GPT, Claude…) and that these AIs will spot the issues in the proposal right off the bat.

“Look at this proposal. Analyze it and tell if it is fair, balanced and spot any potential red flags”

You don’t have to take a prompting course for that!

And I keep wondering: why don’t they review their proposal with GPT, Claude or whatever AI they have in the first place, before emailing me?

Even better, why don’t they add an addendum that tells me that this proposal has been reviewed by Opus 4.7 and scored 8.5 for fairness?

Also, if more and more people behave like me, what is the future of negotiations?

Share

Here are my predictions:

1- Asymmetry is (almost) gone.

Before, an experienced seller or a corporate seller backed by lawyers and people who’ve been in selling for a long time had an advantage. Now, everyone can get almost the same level of expertise pasting the contract in GPT, Claude, Mistral and get:

• it compared with the standard in your industry

• spot anything that is unusual

• and it will explain the kind of situations where you could be in trouble.

The more the AI knows about you, the better and if you’re like me I tell it everything and I organize the information the best I can so that I get maximum support.

So goodbye to, “we know how to sell”. Guess what “I know how to prompt and RAG it better”. Maybe that’s the new form of asymmetry?

2- Transparency exposes imbalance.

When several different models converge on the same critique, it shows that the imbalance is structural in other words, on purpose?

"I sent the same contract to GPT, Claude, and Mistral. All three flagged clause 7. That's not a bias — that's a red flag."

The seller negotiation tactics are now a weakness. More of this and the trust is gone and the next version will be read by a tougher AI, fed with the context of the previous one. You can’t unring the bell and every move is captured and transformed into an asset for the next conversation.

3- Time is the new currency

Sending an unbalanced proposal is expensive.

• it gets rejected immediately

• it damages credibility

• it is … a waste of time

The critical element used to be information (hence information asymmetry) now it is time. What if a competitor sends a better (fairer) proposal shortly?

In our world, sending a high quality proposal, meaning one that the buyer can accept fast, is key.

Time is an interesting beast.

In the old days, in a classic negotiation, time meant a lot:

• answering too fast equates to: “I really need this”;

• silence meant put pressure on the seller,

• slow response usually said “I have other options”.

With AI, this is different. Speed is no longer a weakness, just an indication that the buyer is well equipped and waiting two days to reply has a completely different meaning: If I wait when my AI told me to say no within minutes, it means I chose to wait.

What’s next

So how should you (and I) respond to a proposal today?

Here is where I am and I still have a lot of questions.

1- 100% sure: I use at least two AIs to analyze the proposal. More specifically I do it using GPT and Claude and I do another pass using Claude incognito mode. The incognito mode takes another look at the proposal with no context.

2- this one I am still debating: should I inform the sender that I am using AIs to analyze the proposal and that they should make sure their proposal is fair before sending me a new version?

I’m curious on how you are managing this situation. Please comment !

Subscribe now

In the last year, we've seen an explosion of content, mainly driven by the use of generative AI. Tools like GPT, Claude, Gemini and others create emails, white papers, research articles and I would say most of marketing content we see today.

The result is that we are drowning in content, which can now be produced at a fraction of the time (and cost) it was just three/four years ago.

Yet, this productivity gain has triggered a major crisis of credibility.

When we read an article (like this one), we now question its origin and the author. Is it really written by a human, by ChatGPT, or with a mix of both? For most people this uncertainty is a major barrier to trust.

We thought that technology would save us. However, these hopes are collapsing, Watermarking and AI detection tools are showing serious issues:

• Google SynthID can be erased with simple editing (https://www.lifewire.com/google-synthid-watermark-ai-images-7964120)

• AI detectors are not reliable ( https://hai.stanford.edu/news/ai-detectors-biased-against-non-native-english-writers). Since I am not native, everything I write is probably flagged as A(lien) Intelligence. (documented but still....).

Anyway, the question "Human or machine" sounds so much 2023. The question for today is

"Does it matter who created the content"? 

Hybrid Content Is Here to Stay

AI is now a partner for writing. It's true for most people and it started a long time ago from spelling/grammar checks, to auto complete, to ... Google search. For non native writers like me the transition is natural.

People can use AI to support writing at different stages of the creation process:

• validate ideas (does what I am writing make sense? )

• expand on initial ideas ( any aspects I may have missed in looking at this question)

• check the structure of the article (any better way to organize, rewrite my article)

• check the actionability ( what questions would a reader have after reading this article - this one come from Stuart McFaul )

• does it look "my style" (with a GPT)

• and the usual spellcheck, grammar check...

At the end of the day, the goal is to produce an article that is credible and efficient (I am talking business content, not poetry) and most B2B articles, most white papers are now co-auhored.

So what sets the good content apart?

1. Original Perspective – Does the content offer a viewpoint grounded in experience or strategy?

2. Domain Knowledge – Does it reflect depth, not just surface fluency?

3. Credibility – Does a real person with a reputation stand behind it?

This isn’t just a theory. In a recent interview published by Columbia Journalism Review (https://www.cjr.org/feature-2/how-were-using-ai-tech-gina-chua-nicholas-thompson-emilia-david-zach-seward-millie-tran.php) Emilia David, an AI reporter at VentureBeat, articulated this boundary clearly:

“Writing is hard, and it is my least favorite task, but I do not want AI to write for me… I want my readers to know that I am not just rattling off facts but helping them make informed decisions.”

Trust Isn’t in the Content—It’s in the Network

Marc Meyer captures well this shift in this very interesting article (https://www.linkedin.com/pulse/impending-inflection-point-ai-future-social-media-marc-meyer-v0cge/) : “Synthetic content will become the norm… The future of social media won’t be defined by content creation, but by the ability to discern what’s real and who to trust.”

If we agree that content alone can't carry trust, then what can do it ?

Having worked in social media for so many years, let me offer a solution:

People, Relationships, Social Signals.

What really matters is not whether a sentence or an article was generated by AI, by a human, or by a hybrid. It is whether is was endorsed by someone with expertise, credibility , and... influence.

This is the logic behind peer review in science, citations in academia, and editorial standards in journalism. It’s not just what is said—it’s who’s standing behind it.

Share

Peer Validation at Scale

At eCairn, we have listened to tribes, communities and micro influencers for decades and we asked: What if we could scale peer validation through networks of experts/ key opinion leaders, influencers (all types)?

As an example, with our platform’s audience intelligence engine, we examined the behaviors of over 20,000 AI professionals across research, media, and enterprise. We aren't looking for the most liked or most followed—we are looking for who experts are engaging with, and what kinds of content they shared or debated.

Applying it to this article. (agree it's a little bit of a disturbing recursive) we surfaces these quotes:

• "Trust is no longer a soft value. It’s a monetizable relationship, a strategic asset… Because when anything can be rendered, only trust can be earned.” Lukas N.P. Egger

• " Co-creation with trusted partners (including academic institutions) is emerging as a powerful middle ground.",

Pascal’s Substack

Summary of Day 2 of The Generative AI Summit 2025, London Edition. Generative AI must not be a curiosity—it must drive measurable business value...

Day 2 Summary: Generative AI in the Enterprise…

Read more

a year ago · Pascal Hetzscholdt

• Pascal Hetzscholdt

These weren’t editorial picks. They emerged organically through network validation—the same way trust circulates in any professional community.

Mapping Credibility: The AI Influencer Graph

Our approach is not just about quotes—it’s about structures.

We map social networks among AI thought leaders: who follows who, who collaborates, and who cites which work.

The social graph forms the architecture of trust.

It's not 100% bulletproof.

Keep in mind, people are running campaigns designed for influence and counter influence, plus many KOLs actually work for companies or can be paid to post/like. However if 5-10 people from this group. - all key opinion leaders in AI - have liked, shared or commented on an article .... it has a good chance of being valuable and mostly human engineered, not just a remix of old content.

So this brings me the following questions, depending on which side of the content you stand

• How are you filtering signal from noise in today's AI-saturated content landscape?

• How can you ensure your content earns the trust and amplification of key opinion leaders in your field? What makes content shareworthy to the experts who matter?

Subscribe now

In the beginning, there was standalone Generative AI.

People interacted with tools like ChatGPT—the most popular large language model (LLM)—by crafting prompts and sometimes uploading documents as contextual input. This approach worked remarkably well for many “one-off” tasks knowledge workers faced daily. It’s safe to say that this remains the starting point for the majority of users; even today.

However, since 2023, the landscape has evolved rapidly. Innovations have been introduced to maximize the impact of Generative AI while addressing its limitations and risks.

Compound Generative AI

LLMs, by design, are generic tools. But generic doesn’t always mean optimal. To address specific use cases, a growing number of specialized models have emerged, fine-tuned for particular domains or data types. For instance, Med-PaLM outperforms general-purpose models in healthcare, while Claude excels in text generation and Gemini stands out in image analysis.

Today, users often switch between multiple models based on the task at hand.

My personal toolkit includes GPT, Claude, and Perplexity for daily tasks, while Gemini has proven invaluable for certain client projects.

Subscribe now

In some cases, the requirements of a task are so specific that standard models fall short. This is where fine-tuning enters the picture. While fine-tuning has become less popular in 2025 due to advances in extended context capabilities, it remains a valuable tool when precision is critical.

The evolution of LLMs has been marked by increasing specialization, enabling them to tackle a more diverse array of challenges.

Nevertheless, this is how the LLM layers has evolved over the years to address the diversity of problems to tackle:

Beyond Standalone Prompts

From a productivity standpoint, standalone prompts, while powerful, are not a panacea. Users now need to:

• Reuse and refine prompts.

• Share prompts with teams.

• Develop multi-user applications with complex workflows.

OpenAI’s GPTs, which allowed users to package LLMs, prompts, and documents (in Retrieval-Augmented Generation, or RAG, style), represented an early attempt to address these needs. They didn’t gain significant traction and they are back at it with Tasks.

At the enterprise level, the demands are far greater. Organizations require environments to safely create and deploy a wide range of LLM-powered applications:

1. No-code platforms for business users to build and share simple applications without needing developers.

2. AI studios for more complex tasks, where developers can incorporate additional AI components, such as prompt sanitization or anonymization.

3. Full frameworks for developing robust, enterprise-grade applications, often relying on verticalized or fully packaged solutions.

To address the complexities of scaling Generative AI applications, companies are investing in GenAI MLOps frameworks that manage prompt workflows, monitor hallucination rates, and ensure models stay aligned with enterprise objectives.

The Role of Data and Context

One undeniable truth about LLMs is their propensity to hallucinate—they are, by design, predictive models rather than truth engines. The only way to ensure reliable outputs is to ground their responses in context. And it’s even not guaranteed.

But context has limits. The volume of data an LLM can process in a single interaction is finite (even as this limit grows). To address this, organizations have developed:

1. Databases to store and retrieve relevant content on the fly.

2. Semantic databases and embeddings for efficient information retrieval.

Raw content alone, however, is insufficient for more advanced reasoning.

The next frontier lies in feeding LLMs structured data, such as Knowledge Graphs, which allow for richer, context-aware interactions and a much better strategy to fight hallucinations.

Knowledge Graphs are transformative because they provide LLMs with a structured view of the world. For example, deciding whether an “address field” is represented as street/city/country or as geographic coordinates (latitude/longitude) can significantly impact the model’s precision, privacy compliance, and overall utility.

This is where ontologies come into play, optimized either for retrieval or reasoning depending on the need.

Beyond the Essentials

In addition to data, other critical components are required for a functioning Generative AI ecosystem:

• Security and governance to protect sensitive data and ensure ethical use.

• Integration with existing enterprise systems for seamless workflows.

• Monitoring (both automated and human) to ensure performance remains consistent and aligned with objectives.

and voilà !

Share

This evolution from standalone to compound Generative AI has laid the groundwork for the next stage: Agentic AI, where systems will act autonomously, leveraging diverse data sources and reasoning capabilities to deliver transformative outcomes. Stay tuned for Part 2.

Lately, I’ve been wondering about the implications of using tools like ChatGPT—not just as an assistant but as a potential credibility argument in my sales pitch.

Let me take you through this thought experiment.

The Pitch That Changed the Game

I’m pitching Luxury fashion brands on an opportunity that feels innovative and necessary: focusing on influencers on Substack. The strategy?

1. Listen to what these thousands of influencers are saying.

2. Engage with them meaningfully.

3. Build avatars based on the content they generate.

To refine this pitch, I turned to ChatGPT, feeding it rich data from Substack—content mentioning the brand, content about sustainability, and the intersection of these topics with fashion. I asked it pointed questions:

• Which function within the brand would benefit the most from this approach?

• Could this opportunity address the brand's top three challenges?

The result was a solid pitch, sharp and tailored. But here's where my doubts creep in.

Who Do They Trust More: GPT or Me?

The people I’m selling to aren’t strangers to Generative AI. These are industry leaders deploying corporate AI models within their companies, fully aware of the power and limitations of tools like ChatGPT. And so, I ask myself:

• Do they trust GPT more than a sales rep like me?

• If yes, why not let GPT take center stage? Should I send them the link to its analysis instead of delivering the pitch myself?

What if, instead of traditional sales meetings and demos, I propose a collaborative GPT session? Imagine both buyer and seller sitting together, prompting GPT to evaluate the partnership's potential, guiding the discussion towards a buy/no-buy decision.

Of course, the key starting point would be for both parties to agree on which AI to use, ensuring trust and alignment from the get go.

Is Sales at an Inflection Point?

It feels like sales might be undergoing a seismic shift. The landscape is filled with startups leveraging Generative AI to:

• Hire sales reps

• Train sales reps.

• Support sales reps.

• Even replicate sales reps (https://www.11x.ai/)

But what if Gen AI doesn’t just support sales but

1. Diagnose the problem.

2. Propose tailored solutions.

3. Recommend adjustments—whether it's pricing, features, or strategic fit—to maximize ROI.

The AI wouldn’t just facilitate; it would mediate, offering unbiased, data-driven insights that both parties could trust.

Are Startups Solving Yesterday's Problems?

Here’s the uncomfortable truth: humans are great at solving yesterday’s problems, like extending their home when their kids are gone.

Subscribe now

My Next Move

For now, I’m sticking to the human element, while subtly integrating AI into the process. But I’m tempted to experiment with transparency:

• Sharing GPT’s analysis directly with potential clients.

• Framing the pitch as a joint discovery process rather than a one-sided conversation.

Because if the future of sales is AI-powered collaboration, then the best thing I can do today is embrace that shift—before it makes my role redundant.

Share

( ideas are mine, typos are mine, structure and style are somehow from GPT, although it’s trained on my past “human” written blog posts)

If you have a chance, I highly recommend reading Machines of Loving Grace by Dario Amodei, CEO of Anthropic and creator of Claude: https://darioamodei.com/machines-of-loving-grace.

It is actually very interesting .. and very disturbing.

Machines of Loving Grace (click for the poem)

Amodei claims we’re very close of reaching extremely advanced AI by 2026. This lines up with Sam Altman’s forecast of super intelligence within the next 1,000 days: https://ia.samaltman.com/ . Note that both are avoiding to use AGI (Artificial General Intelligence) wording.

Of course, OpenAI and Anthropic are actively fundraising, so it's part of their founder job to sell their company and tell the world how amazing their work is.

Still, Amodei has always been known for his conservative position on AI—his vision for Anthropic has focused on keeping AI manageable, understandable, and reliable. His current essay, though, hints at a much faster, more powerful transformation than he's suggested in the past.

What I really liked about the essay:

• The images he paints are really powerful to make you understand what it will be like: “Getting 100 years of progress compressed into 5-10 years”, “ Building a country with millions of “people” with Nobel price level intelligence (in all disciplines)”.

• His prospective really shift the focus away from pure intelligence, which will soon be abundant, and toward other real limits we still face: physical resources, experimental timelines (he uses aging research as an example), and the laws of physics themselves.

• The section about biology and neuroscience are very developed. Clearly he has got lots of expertise in these fields.

  Share

For the rest of the essay and more specially the parts about government, politics and democracy, I am not so optimistic.

The assumption that progress is always good seems, to me, naive. Looking back at history—the Roman Empire, European colonialism, Slavery, the Native American genocide, and the Third Reich—when one nation holds a technological advantage, it often starts by using it to destroy its rivals.

In a more recent (tech) history, we already know with social media how technological advancements are used by bad actors. Platforms like Facebook (now Meta) have been implicated in scandals such as Cambridge Analytica, which used data to manipulate electoral outcomes. Twitter has repeatedly been criticized for enabling and amplifying divisive figures (starting with D Trump) and for misinformation, contributing to political polarization. Meanwhile, TikTok and Instagram face ongoing lawsuits concerning privacy breaches, discrimination, and terrible impacts on mental health.

Why should we expect the rise of AI to be different?

What will become of humans … and god?

As for the closing chapter. What will human do when they no longer are needed in most of the production process or progress. I frankly don’t know.

A few AI pioneers come to mind here:

• “If a machine can think, it might think more intelligently than we do, and then where should we be?” — Alan Turing

• “I visualize a time when we will be to robots what dogs are to humans, and I'm rooting for the machines.” — Claude Shannon

So, we should start by treating our dogs kindly. Maybe the future for humans looks more like that of dogs—or even worse, as cattle or horses. At best? Cats !!

Last but not leaset, no mention of what would became of god if a superior AI intelligence is brought to “life”? All religions consider human life as unique and sacred. How will they adapt to a world run by AIs?

Now, I do have a one more question and I don’t frankly know if this is a naive or an “elephant in the room” question.

Why do VC invest in OpenAI, Anthropic if the end game is to make intelligence a commodity?

If powerful AI brings infinite intelligence, the value of intelligence would automatically drop zero (supply and demand).

Companies like Anthropic and OpenAI are nothing else but “intelligence” : engineers with superior abilities, patents, know how … So why would an investor invest in companies whose ultimate goal is to disrupt/destroy the value of the IP/Intelligence that they are supposed to create?

Are they driven by pure philanthropy and the belief in a better world, or do they expect some kind of massive, shared prosperity?

I’d really love to see what that VC pitch looks like.

Subscribe now

More on the same topic/article:

The Algorithmic Bridge

Even If We Win the AI Game, We Still Lose

A blog about AI that's actually about people…

Read more

2 years ago · 36 likes · 9 comments · Alberto Romero

Understanding AI

Six principles for thinking about AI risk

When OpenAI released GPT-4 in March 2023, its surprising capabilities triggered a groundswell of support for AI safety regulation. Dozens of prominent scientists and business leaders signed a statement calling for a six-month pause on AI development. When OpenAI CEO Sam Altman called for a new government agency to license AI models at a…

Read more

2 years ago · 94 likes · 36 comments · Timothy B Lee

• https://www.lesswrong.com/posts/oJQnRDbgSS8i6DwNu/the-agi-entente-delusion

@dominiq

Personal Experience with acculturation training

Last fall, I had the opportunity to develop and deliver an acculturation training on Generative AI for a large company in France.

The participants gave me very good feedback. However, the organizers of the training were not fully satisfied. They were expecting a more "hands-on" training that included prompt engineering workshops and direct applications of GPT for the executives in their day-to-day roles.

It didn’t feel 100% right at the time but I didn’t really know why. So I paused my acculturation training and moved on with consulting and product innovation.

I confess I am not a big fan of prompt engineering. I even think the need for prompt engineering is transient and should disappear as models get better and better but this is a different story. 

The training I designed was more of an introduction to AI and Machine learning, with the goal to make the executives understand how machine learning was different from traditional software, how models were built, how data were selected, prepared and when it comes to generative AI, looking at where it makes sense and where it was a very risky technology to use. I also stressed on hallucinations and bias.

It was clearly not a hands on guide on how to use ChatGPT and improve prompting ability !

Executives using GPT 

Through my consulting work, I meet executives who are interested in integrating AI, particularly generative AI, into their businesses. 

The discussion often begins with me asking about their previous exposure to this technology and a common response they provide is that they are using GPT once every so often and have participated in an "acculturation training." These training usually revolve around using GPT and advanced prompting.

I can tell, after discussing with these executives - and I have talked easily to 80-100 director level and above last year-   that these “hands-on” trainings are not really helpful and can even be counterproductive.

Here are the reasons why:

The Pitfalls of "Hands-On" GPT Training

The first issue I see is that these training usually don’t go back to the core of machine learning which is probabilistic by nature and therefore only correct to a certain extent.  The consequence is that a generative AI can’t really be used in a process that is mission critical or requires predictable and 100% trustable results. 

Its applicability is also very different for people who hold creative jobs ,where an hallucination is an opportunity, versus people who hold a production job ,where an hallucination is a failure.

Here is a great video from Allie Miller (45mn) summarizing the difference between using GPT as a creative or as a productive job.

https://maven.com/p/cc6f1b/how-to-use-ai-to-10x-your-productivity

The second issue is the overconfidence in what an LLM can do.

Executives usually have use the LLM for trivial or content creation use cases. In the real world, a realistic solution often combines different techniques/models to achieve accurate predictions. 

As an example, I just reviewed a platform that includes:  generative graph and text transformers to handle complex data and fill in missing information, XGBoost decision trees to predict outcomes and combine intermediate predictions, Named Entity Recognition (NER) to extract important entities from text, Bayesian optimization to fine-tune model settings, Ensemble learning to improve the prediction performance, and Shapley Additive Explanations (SHAP) to explain how each feature impacts predictions, making the model more transparent and trustworthy. 

This to produce a predictive model that is best of breed (~80% accuracy) but is still not usable if the outcome is mission critical. 

The third issue pertains to the understanding of semantics and the way LLMs models  “semantics”. LLMs generate, summarize, and categorize using “positional semantics,” which relies on the principle that sequences of words frequently seen in the same contexts are similar. 

This approach is the reason why early models struggled with math: in the realm of positional semantics, the vector representation of “seven” is very close to that of “nine”, so in an LLM’s eye, numbers are all mixed up. 

This specific issue has been mitigated with external tools- analytics packages -, better training data, and improved architectures. However, similar challenges persist in many domains and are not easily resolved. For instance, in oncology, the vector representation of one biomarker is often very close to the one of another biomarker. Unless the model is fine-tuned with domain-specific data or unless it is provided with precisely defined context and ontologies, LLMs will mix up things and generate incorrect answers.

The fourth issue involves broader concerns that must be addressed to ensure responsible and effective deployment of generative AI. It’s all the domain of Ethical AI. 

Generative AI models perpetuate or amplify (historical) biases present in their training data , necessitating robust bias detection and mitigation strategies, including generation of synthetic data.  It may not seem critical but if, as an example, marketing material keeps assigning people from minorities in similar roles, this could be very damageable for the brand mid term.

The use of generative AI also often involves processing large amounts of data, raising significant data privacy and security concerns. Emphasizing data anonymization, secure data handling practices, and compliance with data protection regulations like GDPR/CCPA is crucial. Training programs should therefore touch base on regulations and best practices for maintaining compliance.

Last but not least, these technologies are often presented as a new  “software”.

When presented with Gen AI, people “think”: how can I improve my software, my process using these capabilities? This, in my opinion, limits the imagination and hides what is really possible.

I prefer looking at it with the analogy of “let’s say you can hire a junior MBA with an infinite amount of time” that can look at your data, problems and extend it with anything available on the internet. What would they do? 

When you look at it this way, many new use cases come to light: in BI, you can read the web real time and create metadata about .. anything. In software, you can radically change the UX, and even disintermediate most of application software. You don’t have to use the application anymore, the AI assistant is doing it for you.

So I prefer the junior MBA lens. I even wonder whether Gen AI should be managed by … HR and not IT. Maybe not in 2024, but I bet this will be a key question in the years to come. 

Share

Moving Forward with AI Training

Given these insights, it's clear that AI training needs a balanced approach. While practical, hands-on experience is valuable, it is clearly not enough and can’t equip executives to make decisions for generative AI investments.

Executives need a strong foundational understanding of AI principles and limitations. 

Executives must learn to discern where AI can add value , where traditional methods might still be more effective and where it’s a big no-go.

If you're looking for comprehensive training that goes beyond hands-on experience and equips you to make strategic decisions with AI, consider reaching out to our team.

Good morning.

I have been traveling for the last few weeks and as I followed from a distance the news in generative AI (GPT-o, Gemini …) and in politics (Trump saga, rise of the far right in Europe) I couldn’t help thinking about things that do not make sense for me.

I wanted to share with you what I call the “AI paradox” with hopes that you can help me sort this out. Don’t forget to comment !

Where do I start….

With ChatGPT, but to be fair, since the beginning of Artificial Intelligence (AI), people are insistent that AI has to be built for good. Humans are scared of AI going out of control, of behavior that could harm humans or even cause the end of humanity. This fear is present in most science fiction movies and novels.

It made it back to front page last year with controversial discussions between the three Turing prize winners ( LeCun, Bengio, Hinton) on whether we should stop AI research altogether (https://venturebeat.com/ai/ai-pioneers-hinton-ng-lecun-bengio-amp-up-x-risk-debate/)

It’s also the underlying issue of a new Netflix movie.

The critical issues revolve around ensuring that AI systems exhibit "positive" behaviors and thoughts.

Example of these “positive” behaviors include not harming humanity, not being racist, promoting diversity , eliminating gender biases….

The AI Dilemma

Practically speaking, the challenge translates into designing algorithms and systems that do not perpetuate historical biases.

Researchers and developers strive to create AI that can transcend these biases, aiming for a more equitable and just application of technology.

This task involves rigorous scrutiny of the data used to train AI (the training corpus) , as these data sets always reflect historical prejudices and inequalities.

Developers even go into making up datasets !! (synthetic data) that reflect a corrected version of history.

As an example, if you want to correct gender biases, you can preprocess any content with a gender and generate the same content with a different gender, thus balancing the training set i.e

• The doctor just called me. He was very nice

build an alternative sentence:

• The doctor just called me. She was very nice.

and also

The doctor just called me. They were very nice.

add this to the training set and voila, you get a dataset that is better than reality.

The goal is clear: AI should embody principles of fairness, promote diversity, and be free from the biases that have plagued human societies for centuries. This is a noble and necessary objective, considering the influence AI can have in various domains, from hiring practices to law enforcement.

What is crazy to me is the following:

We are acutely aware that these biases are deeply rooted in human history and acknowledge their detrimental impacts. Yet, when we transpose this reasoning to human education, and human behavior, the reaction is often one of fierce resistance.

Discrepancies Between AI Ethics and Human Policies

1. Respect for Human Life: This is the first law of robotics from Asimov which is frequently quoted (https://en.wikipedia.org/wiki/Three_Laws_of_Robotics).

   AI ethics dictate that AI should not cause harm to humans, aligning with principles of safety and respect for human life. In contrast, human societies frequently witness violations of this principle through wars, violence and systemic injustices. We, as a society are clearly not holding human actions to the same ethical standards we set for AI.

2. Fairness and Non-Discrimination: AI is designed to be unbiased and fair, with strict guidelines to avoid discrimination. However, in education and justice, implementing non-discriminatory practices faces significant challenges (recent example with Florida: https://www.edweek.org/policy-politics/whats-with-all-the-education-news-out-of-florida-a-recap-of-education-policy-decisions/2023/08) . Policies like affirmative action are often contentious, and systemic biases in judicial systems persist despite efforts to promote fairness.

3. Diversity and Gender Equality: AI is often designed to promote diversity and gender equality rigorously. In human systems, achieving these ideals is a major challenge. Religions (all of them) from the beginning of time have ensured dominance of men over women and they are still unquestioned. Even in “modern” laic societies, policies aimed at reducing gender and racial disparities and biases persist due to deep-rooted societal norms and resistance to change.

4. Transparency/ Truthfulness in AI vs. Human Systems: AI guidelines emphasize transparency to ensure ethical compliance. However, since the critics of Kant, a hardliner supporter of “truth whatever it takes” we pretty much admit (cf Benjamin Constant) that lying is sometimes a good strategy and a condition of survival and well being for humanity. Should an AI lie to someone asking about information it needs to commit a crime? Should it be transparent when providing a very bad medical diagnosis ” ? I often wish my scale would lie to me ;-)

5. Accountability: AI ethics often stress accountability, expecting AI systems to be auditable and traceable. In politics, lobbies from all kinds are at work to influence decisions. Every government also utilizes underground and non public operations. Whistle blowers are risking their lives for providing us with transparency.

I stop there but the list could go on and on.

The Human Paradox

This double standards reveals a profound paradox: humans are capable of distinguishing between right and wrong, as evidenced by the standards we set for AI.

This double standard has practical implications when it comes to setting the acceptance bar for AI systems. Let’s take autonomous cars.

We know that humans can be terrible drivers, especially when drunk. Metrics show that self-driving cars are already 6.7 times less likely to injure passengers compared to human drivers (source: The Verge). Despite this, any accident involving a self-driving car is seen as a catastrophe, and we demand 100% safety from these vehicles.

I understand that the unknown (we are just starting to use widespread AI in general public) may justify higher standards for AI than for humans. Still we have no experience of an harmful AI whereas we have tons of experiences of harmful humans and we don’t seem to learn much about them.

Conclusion

What are we really looking for in an AI? A god-like figure? A better version of ourselves? A human copy-cat?

What do we mean by alignment? We know perfection is not a workable agenda and the real world needs errors and lies to properly function.

How do we define trustworthiness and ethics? Are we referring to the theoretical version of trust and ethics or a “relative” human version that is open to mistakes, lies, and manipulation?

I don’t know.

I think that reflecting on these questions can help us navigate the complex connections between AI and human ethics, maybe help us finding a more balanced and realistic approach to AI development and deployment.

I also hope that setting the rules for AI will help us question the low bar that we have set for humans and human policies.

Since eCairn's establishment in 2006, we have been at the forefront of the social listening industry, witnessing its transformation from the beginning of social networks to the current era of advanced analytical tools.

Facebook was born in 2004, Twitter in 2006. Linkedin at the time was primarily serving as a repository of professional profiles. Back in 2006, blogs were the primary platforms for online discourse.

We’ve seen the rise of Radian6 (created in 2006) and its acquisition by Salesforce and the success of a few others. Yet, the path for most has been full of challenges.

It was and still is clearly extremely difficult to create a sustainable business model in a field where the perceived value is limited to a select group of experts within organizations- specifically, the social media marketing specialists.

Traditional Social Listening: A Labor of Insights

Typically, social listening within corporations is done as follows:

• A dedicated social intelligence team, usually operating under the marketing umbrella, is responsible for delivering insights and analytics, utilizing platforms such as Talkwalker, Synthesio, … and eCairn to enrich the company's understanding and engagement with digital communities.

• This team generates a set of standard reports including metrics like brand share of voice , trending topics and competitive positioning, which often fall short of catering to the unique needs of end-users.

• It also performs ad-hoc studies to research specific products, services, events, … and customer segments

• The visibility and usefulness of social marketing is mainly shown during crisis when management is desperately trying to respond to a story that has gone viral.

When business is good, things are OK. When business is not as good it is the first organization that faces budget constraints.

The results is that nobody is satisfied

• Management perceives little ROI if any from its social listening investment.

• The outcome of the social listening is “one size fits all” and usually the team doesn’t have the responsiveness to provide intelligence when it is needed

• For the social listening team, it's a perpetual cycle of frustration, stemming from unacknowledged value and insufficient resources to deliver.

• Despite their utilities and unique strength ( eCairn specializes in monitoring and profiling opinion leaders within specific vertical) , the social intelligence platforms are often perceived as non-critical, with its value appreciated by only a limited audience within organizations. Vendors can’t build a sustainable business and their users are laid off every few years.

We (at eCairn) often joked that we have more clients than users, as our users introduces us to their new employers, each time they lost their job and moved to another company. A CEO of a company that pivoted from social media marketing to CRM told me a while back: “ social marketing and social selling is something that everyone says is a no brainer … but never implement”.

The Generative AI Revolution

I believe generative AI is making applications obsolete.

All applications.

This is even more relevant for applications whose core purpose is the make data actionable.

Why do you need an application when an intelligence agent (think of a young MBA assistant) can synthesize data for you and provide a recommendation?

After all, when using Google Maps, when you’re told to turn left at the curb … you turn left and don’t ask for reports and datapoints justifying that going right would be a bad decision.

Social listening is mainly reading and interpreting data and generating a description/ summary of this data to someone with a particular profile/need.

• generating a description/ summary : This is the core feature of generative AI.

• reading and interpreting data : This is what retrieval augmented generation (RAG) offers.

• someone with a particular profile/need : This is what prompt engineering delivers.

  

This has profond consequences for every player down the chain

For users, the integration of generative AI makes social listening obsolete.

Imagine an enterprise equipped with a conversational agent capable of reading and analyzing social media data in real-time. This AI assistant could be accessible to all managers and employees, enabling them to query specific research questions and receive instant responses.

There are many benefits:

1. Immediate answers

2. User friendly conversational interface (like ChatGPT)

3. Ability to tailor the analysis to the user profile and refine it thru ongoing questions/conversations

4. Understanding the landscape beyond the capability of a single tool drawing in diverse data sources

5. Cost savings

For the social marketing people.

Most of them will lose their jobs. Humans will still be required to align the system to the company’s objectives, fight hallucinations and improve the system overtime. Human will also have to ensure the quality/availability of the data sources and personalize the answers with prompt engineering.

However, this is a challenge for many people in the field as these activities require different skill sets compared to what social marketers have been trained on.

For vendors like eCairn

Most of the companies will die and close door. The only pivot that is available for them is to become a specialized engine generating genuine data i.e

• data that you can’t buy off the shelf or get from social media platforms (Twitter, Facebook, Instagram…)

• nor that you can construct easily using a vanilla generative AI.

Fortunately for eCairn, generative AI does not (yet?) understand the multi facet and nuance of influence and the subjective nature of tribes and communities.

So the next step for us is to carefully curate data, educate/train the generative AI to fill that gap and become the Influencer/Tribe- GPT.

Subscribe now

Fig. 1- The Long and Winding Road

AI research has experienced two winters already, where unreasonably high expectations paired with disappointing deliveries were met with skepticism - even derision in some quarters.  These AI winters were inflection points in the history of AI research that saw funding largely vanish, and research opportunities wither on the vine.  Each AI winter, through the efforts of dedicated researchers, was (eventually) followed by a resurgence in the field.  We see that many technologies advance in such a boost-and-burst fashion, driven by hype, fear, and greed - those typical cycles are often called ‘manias’.

CURRENT BOOST CYCLE - GAI AND LLM

The current ‘boost’ cycle for AI research is characterized by high, almost manic, interest in Large Language Models (LLMs) and Generative AI (GAI).  We recognise this as a possible point of inflection from the epoch of AI technologies being used predominantly to gather, process, and synthesize huge volumes of data, into a new epoch where AI generates, rather than just collects and processes, text, images, videos, and audio at an unprecedented scale, speed, and scope

As the generative capabilities of AI are streaming torrents of data onto the internet, concerns are being raised about the social, economic, and political impacts of those systems.  We risk unchecked and unfiltered AI-generated data becoming the predominant source material for the very AI systems that are generating more data to flood the internet.  A valid concern is that we may be embarking on a path that may allow AI systems to own our narrative and write our history.

GAI and LLMs have the potential to impact the daily lives of a vast number of people worldwide.  Indeed, we already see that the general community has both an awareness of, and trepidation about, the capabilities, and unfettered use, of LLMs.  We believe that governments are sensing the concerns of the public, and of experts in the field, and legislation, regulation, and certification of GAI in particular, and AI generally, is inevitable.  We believe strongly that AI researchers should be part of that discussion.

LIFE ON EARTH AND INTELLIGENCE

The rise of huge infrastructures, running gigantic Artificial Neural Network (ANN) models, will likely enable some new/novel developments in the field.  While the new and novel developments will no doubt have a beneficial impact on society, these installations will have the potential to impact people and the environment negatively - consuming vast amounts of energy and generating huge amounts of heat and other waste products.

As we already observed, technologies often advance in a boost-and-burst way, with each ‘burst’ usually having its own, unique, impetus. Two world wars in the previous century advanced nuclear, electronics, and computer technologies, and, by way of analogy, we expect that current and future conflicts will see the rise and importance of military drones - ground-based, airborne, marine, and submarine.  By some indications, drones and autonomous vehicles might play a key role, just as cannons and tanks did in previous conflicts.  Embodied AI - drones - might emerge as the key military/war technology of the 21st century - yet another argument that we are seeing an inflection point in AI advances.

Fig. 2. Life on Earth and Five Layers of Intelligence

We think that AI may morph into a general Science of Intelligence - as a multidisciplinary field focused not only on mimicking, or even surpassing, human behavior and capabilities, but also on the other layers and kinds of intelligence that enable life on Earth. It will take a very long time to see this field established as a science, and consequently AI research will largely continue to be about engineering systems that (try to) mimic human intelligence, while the Science of Intelligence would frame all life on Earth (Fig. 2) as the consequence of variety types of interacting intelligences - of plants, insects, animals, humans, etc. - with the new science eventually underpinning far-reaching AI research and development.

INFLECTION POINT ?

The relatively recent and rapid rise of, and hype generated by, LLMs in general, and ChatGPT in particular, has caused interest in AI to spill over from academia and specific interest groups to the wider public. Businesses are embracing AI; whole industries are changing to accommodate AI; the trajectories of economies are changing because of AI. But it is not just businesses and economies that are changing.  Advances in computer hardware and AI technology have seen the increased use of military drones in battlefield situations, thus changing the nature of conflict and affecting millions of people. The world will change more rapidly than most people imagined because of the rapid rise of GAI.

Fig. 3 - From Data Aggregation/Analytics  to Content Generation/Synthesis 

During the long history of AI research, the initial focus in the Embryonic phase was largely on a programmatic approach, which evolved into a training approach, often characterized by Artificial Neural Networks, as we moved into the Embedded AI phase. With the advent of GAI and LLMs, the focus of AI research has shifted from the creation of analytical artifacts (e.g. predictions, diagnosis, synthesis, etc.) into massive content generation (e.g. text, image, music, video, code, etc.). We see this as an inflection point in the ongoing history of AI research (Fig. 3).  

We believe the current high level of interest in Artificial Intelligence, evidenced by the implementation of national programs by governments around the world, huge investments from industry in terms of money and time, increased marketing and advertising of AI-enabled devices, and the chatter on social media from the general public, is good evidence that we are at, or approaching, an inflection point in the development, and crucially, acceptance, of AI systems.

Since the introduction and massive success of ChatGPT, an increasing number of companies have embarked on creating  their own Corporate GPT, either fine tuning existing models or using retrieval augmented generation (RAG)  technique  to enrich foundation models with their proprietary content.

The rise of Corporate GPTs

The trend started with BloombergGPT (https://arxiv.org/abs/2303.17564), which built a large model incorporating financial Data in March 2023. Since then most companies have experimented with GPT technology (or alike)  crafting their own version of MyCorporateGPT. A friend of mine even mentioned a  PastaGPT developed  by a food company !!

In my consultations with brands in the US and in France, it is usually one of the first requirements that comes up:

“we want our employees to use our own version of ChatGPT”.

Such projects are excellent for experimentation. They are also a good fit for specific use cases.

As an example, a corporation may want its public facing communication people to reflect the values, the wording and style that it has spent time to establish, document and enforce.

However, the shift towards universal adoption of "Corporate GPT" within organizations comes with its set of challenges:

1. Historical Data Bias: Relying on corporate archives to train these models can lead to outputs that mirror past business practices, essentially echoing the mindset of employees and management from a few years before.

2. Echo Chamber & Lack of Diversity: A model trained solely on company data may cement existing viewpoints, stifling innovation. This is particularly risky if the company overlooks emerging trends or critical technologies, as the model would perpetuate the "business as usual" approach over fresh, innovative solutions.

3. Impact on some Employees' Roles: For employees tasked with market research, technological scouting, or driving product and strategy innovations, relying on an internal model could hinder their ability to incorporate external insights and learnings.

4. Misalignment with Company Culture: The approach to knowledge sharing and access within Corporate GPTs must resonate with the company's cultural values. Providing too little information can lead to employees not using it. Exposing them to information that management prefers to remain confidential can bring all sort of trust questions & issues.

Before diving into the development of a Corporate GPT, companies should closely assess the needs and roles of their employees.

For instance,  customer service representatives in call centers could benefit from a GPT integrated with  internal product documentation and past interactions, enhancing their ability to provide better responses. Similarly, for employees engaged in technology scouting, their GPT should be a gateway to external publications, expert opinions, and technological discourse in their field.

One size does not fit all.

The reality is that different jobs will require different “assistants” and therefore different models and different datasets to personalize the model to a task, a role and a function.

This will create a nightmare for companies who still have a disorganized knowledge management system and poorly documented organization, processes and guidelines.

I would argue that each and every role in a company requires its own model.  Sure there is a common body of knowledge that is beneficial for everyone, starting with org charts, employee communication … but to really benefit from a useful AI assistant, specific needs arise:

• An HR professional would need to access company HR policies, labor laws and regulations, industry benefits, compensation and salary structures, competitive information on hiring, pay …

• A procurement employee would need supplier management strategies, access to ERP, history of supplier evaluation along with import/export regulations, vendor contracts, purchase history ….

• A marketer would need results from market research, ad performance data, industry specific trends and ideally ;-) insights from influencers and key opinion leaders in the domain ( a little promotion there: www.ecairn.com) …

• A customer support representative would need to access product and services help and FAQs, call logs, access to CRM …

For each role, a blend of corporate knowledge, common to all employees and specific knowledge to the role or even to the task would  be required.

This brings a strategic challenge relative to who has access to what information within a corporation. 

Beyond initial pilots, deploying assistants within an enterprise will require a program, policies, an organisation and a technical infrastructure.

Subscribe now

Addressing knowledge access

Determining who has access to what information within a corporation is strategic.

This question is a prerequisite for any Corporate GPT / Enterprise GPT initiative and the answer to this question has a profound impact on company culture. 

Some companies may opt for a compartmentalized knowledge structure, some may be more open.

There may even be a trade off between the company and the employees. I remember deploying a knowledge management platform in a large corporation's sales department. We struggled to convince sales reps to share their prospect insights. Why would a seasoned salesperson share their valuable contacts, which directly contribute to their success and compensation? The exact same questions arise when creating a generative AI assistant.

Hence,  although Corporate GPT makes a lot of sense on the surface, initiating projects with well-defined and bounded tasks or functions seems a more pragmatic approach … except for companies like Bloomberg who have assembled a unique body of knowledge that can be monetized directly through a model.

Share

No, the only answer I want to get is the one that resonates with every french soccer aficionado : “les 3 points”. ( the three points is what the team scores when winning the game)

“L’important c’est les 3 points ” is an iconic opener for the After foot radio show and a close to cult mantra.

Distributional semantics

You might be wondering?: what does it have to do with LLMs, ChatGPT and the broader landscape of generative AI ?

At the core of LLMs is this idea that words are defined by their semantic and that semantic can be mathematically coded using the principal of distributional structure ( pioneered by Zellig Harris in 1954).

By analyzing vast amounts of text data, these models learn to associate words with similar meanings based on their contextual usage.

To make it simple, imagine the internet with a large number of sentences :

1. "The dog eats its kibbles."

2. "The cat eats its kibbles."

3. "I am taking the dog to the vet."

4. "I am taking the cat to the vet."

5. "My favorite pet is dogs."

6. "My favorite pet is cats."

In this massive collection of sentences, “dog” and “cat” are frequently used in the same context, therefore they are “semantically close” words and they belong to the same category.

This is how LLMs and ChatGPT understand language.

Symbolic vs Semantic vs Tribal

Where am I going with all this?

My point is that language is more than semantic.

First, there is a symbolic aspect of language. Symbolic language involves layers of meaning that go beyond the literal, often tied to cultural or historical contexts. For instance, the color red is associated to revolution for many. Similarly, sentences about a black swan (cygne noir) and “un aigle noir” have nothing to do with birds. GPT does a reasonably good job at capturing this symbolic aspect. GPT (click) got the “cygne noir” part correctly.

But beyond the symbolic, there is another facet of language which I define as tribal.1

Humans do use language to create a sense of belonging and identity. They use it as a mean to associate themselves to groups & tribes and differentiate themselves to the rest of humanity.

Sometimes people do this to resist (minorities and people under oppression are forced to invent their own code and language), more often they do this to exist…

The objective of language is, in that case very similar to a deliberate choice of clothing or style.

emo, sporty, boho

The same way people would create and adopt a dress code, they invent a particular slang sometimes going beyond lexical distorsion and changing the syntax.

Subscribe now

Why is it an issue for LLMs

The challenge for LLMs in capturing the "tribal" aspect of language is significant. The vast and diverse nature of human language, coupled with the pace at which communities innovate and adapt their linguistic expressions and slang makes it close to impossible for these models to keep up.

"Humans will always alter language faster than AI will comprehend it"  (own citation)

LLMs are trained on an immense corpus of data: the internet and other corpuses. They kind of summarize it in a low resolution version. They may expand to trillions of parameters, it’s not gonna be enough to capture the finesse of the millions of “slangs” out there.

Already the French language is underrepresented on the net and represent only about 5% of the corpus.

So, if you take the “french soccer fan community” (~ 14 million people listening monthly to this podcast). These millions soccer fans rarely write more than a tweet/insta post every once in a while …. there is no way any model, whatever the size would get the granularity to capture the finesse of the “after foot fans slang”. The volume of content produced and contributed to LLMs is just a drop in the internet.

The implications are much wider than soccer! [yeah that’s just a game … they say].

I recently worked with a retailer for teenagers and it’s a major challenge for them to build a chatbot that respond with the right tone, a model that create marketing messages that are cool, trendy and current with the fast-paced world of teen fashion.

Glimpse of hope, Social data

Although the picture is gloomy, there are options to capture the lingua of tribes.

Our strategy (at eCairn) is to source content right from the social media platforms.

We do it is by mapping the influencers ( thousands of them) and content creators in a tribe, collecting the content they write and enriching models using RAG (Content Injection) and fine-tuning techniques.

Fashion influencers on substack

We will elaborate this technique in a subsequent post. However if you’re interested in how we do this, just contact us!

Conclusion

ChatGPT is relatively new and even if millions of people are using it, we, as a society have not yet adapted to a world where a lot (the majority?) of content is generated by AI.

I bet that humans (us) will not be satisfied by this new content/information landscape and we will react.

We need linguistic diversity. We speak many subculture languages and we want to continue doing so.

I bet we will increase the pace at which we build slangs and tribal languages not only to set up apart from the mass but also to differentiate us from AI.

The future is linguistic diversity.

Just sharing a short experiment with ChatGPT.

I am in the middle of developing a training package on AI/ Generative AI, something simple for non-tech people. The challenge is to convey complex ideas in a simple manner for a non-technical audience.

As part of this training, I wanted to create a slide that shows a very simple linear regression along with how to measurement error.

So I started "googling" looking for images and this is what I got:

This was exactly what I was looking for, except that all the images in these articles are copyrighted and I could not use them in a training material that will probably be posted on the web (been there before and had to pay some $$ to Getty).

What were my options?

• spend time looking at all these sites and see whether I could use a “free from copyright image”

• create the slide from scratch. I’m not that proficient on powerpoint so it would take me easy 1-2 hours to make something nice… and correct!

None of this is very exciting. So I decided to look at it differently.

I went on a real estate portal, collected 15 listings of houses on the market in Grenoble along with the surface an price and I asked ChatGPT to perform the linear regression and plot the outcome.

Bingo, it took me 10mn to do that and the result is perfect

The result? In just 10 minutes, I had a custom-made, accurate, and copyright-free illustration of linear regression.

But why stop there? To add a dynamic element, I asked ChatGPT to generate an animation demonstrating the progressive improvement of the regression line. This visual was intended to simulate the concept of 'learning' within the AI model.

[ I know that for the data science purists, this is a simplification—after all, in reality, the optimal solution is often computed in one step via the "normal equation."]

Despite this, the value of the animation is really good. It provided a clear way to visualize how a model adjusts to data over time, an essential aspect of machine learning.

Nevertheless, I find it amazing that it was faster and much more interesting to build a linear model than work on powerpoint.

ChatGPT is really amazing.

When I think about LLMs, what keeps me awake at night is not the existential threat on humanity, nor the significant toll that Generative AI will have on our planet resources. It is even less the question about regulation or the right claims for copyrights.

No. It is the more pervasive influence of LLMs on how we communicate, think and establish norms within our culture.

LLM shapes opinion, wins elections and sets cultural norms … and there is not much that we can do about it

But instead of going into the whys and hows, I will share with you two examples from this week recent news.

Let’s start with the title of this post:

Should the LLM produce inclusive content?

Last week in France, President Macron made a major speech stating that there is no need to bend the French language to enable inclusive writing.

For non french speaking people, we don’t have the (singular) “they” in french for gender neutrality. So you translate: “Someone just broke into my home, they only took my wallet” in “Quelqu'un vient de s'introduire chez moi, il n'a pris que mon portefeuille.” Il is He.

This is quite a hot topic in France. It was also discussed in the French parliament on the same day.

Taking into account that all major LLMs are created and “censured” (aka guard-railed) by companies and their employees from the San Francisco Bay Area . What are the odds that these models generate more and more inclusive content?

In the blue states where they operate, it is actually the law: California, Oregon, Washington and several other states are champions of inclusivity and have already put “inclusive education” and “inclusive writing” into law.

Granted this will be a big issue for some conservative-leaning “Red” states. Here there are movements to ban books featuring LGBTQ2SIA+ characters…

Here is an example or how generative AI could influence writing: https://chat.openai.com/share/6bd8f326-2898-4b56-922e-cd270d8655f4 (Texts are very poor; and it’s even worse in the inclusive version)

Imagine, over years, students are exposed to an LLM promoting a term like "iel" (the gender-neutral pronoun in French). This term could potentially become normalized, effectively changing the French language—and by extension, French culture—without any policy change or public debate. It's a subtle evolution, one that could happen so gradually it goes almost unnoticed until it's the new standard.

I am personally fine with this ... but my neighbors in Idaho and more importantly most of the french presidential candidates would call this brainwashing.

And let’s say France decides to regulate, will it ask the “Academy Française” to provide real time recommendations to LLM creators and perform quality checks ?

Let’s take another example

How should LLM talk about the Israel/Gaza event?

On another front, consider the challenge of how LLMs address contentious historical events, such as the Israel/Gaza conflict.

I asked (in french) a very neutral prompt to ask to generate a paragraph about the October 7th events.

Here is what Bard produced: https://g.co/bard/share/3abb7bf72649 .

You have probably noticed the level of tension on whether or not to name Hamas a terrorist organization and whether to present Israel's military response as a genocide.

The Google Bard response takes a side: Hamas = terrorist, Israel response is "escalade de la violence" (which is clearly not a definition of a genocide). Period.

The manner in which LLMs present such information is crucial because it can influence public perception and sentiment.

Can this be regulated? It's doubtful.

Regulation would require an overreach into how companies like Google curate and prioritize information, an action that is antithetical to the principles of many democratic societies. (this sentence was written by ChatGPT, not bad !!).

It would be also extremely expensive. A little known fact is that Google employs ~15000 (underpaid) ad quality raters in the US . These native people check the quality of its search results in different languages.

By the way this example also illustrates an important challenge from a technology prospective.

The GPT model (GPT4) was trained from data that pre-date the October 7th attack. This attack clearly changed the perspective that (most) people have on Hamas. How can LLM be updated to reflect this shift without going through a complete retraining?

Additionally, in the case of Bard, it stays updated with the latest information by running a Google search based on the prompt to gather relevant context. Modifying Bard's responses would, therefore, necessitate influencing Google's search algorithms. While countries like China and Iran exercise control over internet search engines within their borders, such a level of regulation is not feasible in the United States or the European Union.

In a matter of months, these models will be everywhere: in our cell phones, Alexa speakers, and even surveillance systems. They will find their ways in algorithms that influence who gets hired, who gets a bank loan, maybe even who is put on a list of suspects. Yet there is no prospect for any kind of a democratic oversight.

Could be worse, these models could come from Moscow or North Korea.

@dominiq

I am sharing with you an experiment we are doing at eCairn, to deliver agents with explicit preference.

Why this? where all solutions out there are developing techniques and strategies to fight biais, there are actually use cases where biais/preference is useful.

As an example, let’s say I am building a conversational agent for Bernie Sanders campaign team (@BernieSanders) . These people would need rapid access to information that aligns with the view point of their candidate. For them, an extremely preference solution is the way to go.

I would even argue that there is benefits in “transparent biais”. What is not OK is when people pretend to provide “objective information” and sneak their own prospectives in the results.

So with this in mind, at eCairn, we have embarked on a journey to build

“ Agents with explicit preference”.

How do we do that?

• First, we gather 10000’s of individuals and organizations in social media who identify with a specific “tribe/community” - this is eCairn’s secret sauce -.

• These groups of people could be “hard core progressives”, “people from a specific city”, “soccer moms”, “experts in data science”, “affluent women entrepreneur in Upper West Side New York”…

• We leverage this group of people as an echo chamber and capture both the articles they like/share (a lot) and their opinions (their own content).

• Much like Bard, or the newly enhanced ChatGPT with search, we scan the collection of automatically curated “hard core progressive opinions and respond to the prompt using the best articles and opinions as context.

Demo

But let’s do a demo, using a controversial prompt (and very relevant in Florida!):

Should book with LGBTQ characters be banned from classrooms?

We ask Bard, GPT and our “Progressive_GPT” the same question for comparison.

Here is the link to the full response: https://g.co/bard/share/f557bec68be2

Here is a link to the full response: https://chat.openai.com/share/84a4c02b-27c8-4ba6-9650-a2f2f23151df

Progressive_GPT

Now let’s look at the Progressive_GPT, powered by eCairn

Let’s now look at how the system works when fed with opinions are articles shared by 1000’s of progressive opinion leaders. Here are a few influential individuals:

To power the conversational agent, we collected ~5000 progressive voices on Mastodon ( we also use Twitter but Twitter is getting expensive since Elon took over, and our findings is that the political conversations on Mastodon are actually richer than on Twitter).

As you can read, the answer is way more direct and affirmative.

Using a similar technique, it is possible to create a myriad of personas… as long as there are sizable groups in social media that identify with the persona.

Side note

The responses from GPT and Bard are supposed to be neutral but are they truly? and who defines the standard for a neutral answer?

For certain questions, refraining from taking a position is far from being neutral.

Share

Share

I recently read this excellent blog post by George Colony: The GenAI Imperative , that clearly underscores three pivotal facets of generative AI: End of the Web, death of Google, the renewed importance of trust.

Although I agree with these observations, I think he is missing the potential of LLMs and embeddings and to the existing list of three key points, I propose adding a fourth: The Power of Document Intelligence.

LLMs and embeddings will unlock an incredible amount of insights (granular insights), that have until now been impossible to discover and utilize at scale.

Embeddings

Let’s look at what embeddings are in more details. In NLP ( natural language processing), embeddings are mathematical representations of words, phrases, or even entire documents in a continuous vector space. This means that each word or document is transformed into a multi-dimensional point, where semantically similar words or documents are located closer to each other in this space.

These words can even be combined and you may build equations like: Queen+ Man - Women = King, or Paris+Spain = Madrid.

from: https://medium.com/@st3llasia/graph-embedding-techniques-7d5386c88c5

But starting with the blog post:

George Colony’s blog post

What I really liked about this article isn't just the elements it emphasizes but also the aspects it omits.

First and foremost, Colony refrains from making grandiose claims about Large Language Models (LLMs) being sentient or getting close to Artificial General Intelligence (AGI) – an intelligence that rivals or surpasses human cognition.

This contrasts sharply with the stories that companies like Microsoft have been promoting: “we believe that it could reasonably be viewed as an early (yet still incomplete) version of an artificial general intelligence (AGI) system”

Moreover, Colony doesn't scare us, cautioning against AI investments due to AGI risks and the end of the world.

Also he does not predict the end of labor and the massive layoffs like Goldman Sachs, predicting 300M jobs lost or degraded!

Instead, Colony offers a grounded perspective on the capabilities and implications of LLMs and Generative AI:

1. A Language-Rich Interface: LLMs usher in an era where individuals can engage in dynamic conversations with documents.

2. Redefining Search: The traditional search paradigm is evolving, and LLMs are at the forefront of this transformation.

3. The Imperative of Trust: With the advent of "context injection" in LLMs, the source and traceability of content are becoming critical.

Getting more intelligence from text

While Colony's insights are invaluable, I believe there's a fourth dimension he hasn't touched upon.

The essence of LLMs is intrinsically textual and for me the #1 benefit of LLMs (and embeddings) is to overlay text with a representation that enables connection, abstraction and manipulation at semantic level.

4. The Power of document intelligence: LLMs, with their ability to process fine-grain embeddings, will revolutionize business intelligence. They will empower us to extract knowledge from documents at a semantic level and learn from combining documents, comparing documents, predicting based on meta data …

To illustrate, here are a few examples:

Retail

Imagine a retail store enriching its customer database with insights from consumer locations.

By aggregating articles about various cities, the retailer can amass a huge amount of metadata about these cities.

This could range from conventional metrics like population density, average age of the population, average wealth … to nuanced attributes like, who knows: the city's proximity to a river, insights about the weather, the presence of a library, a swimming pool, sentiment about the housing market, local preferences from a specific category or even the prevailing political views.

Using LLMs, such knowledge can seamlessly integrate with existing business intelligence frameworks, offering unprecedented insights.

Maybe the retailer can correlate its books sale with these parameters and make rich/granular predictions.

HR

The vast majority of data handled in HR is textual.

For instance, consider the hiring process in HR.

Traditionally, HR matched keywords, potentially overlooking very good candidates with synonymous but not identical skills or job titles.

Embeddings in HR are game-changers, linking resumes, job postings, skills, and company values with unprecedented precision. They identify subtle skill correlations, bridging keyword gaps and making sure no promising candidate is missed.

Furthermore, embeddings could help evaluate cultural alignment by encoding company values into vectors and measuring the similarity between a candidate's values and those of the company.

Scope can also be extended to career plans, evaluations, organizational descriptions, company culture, company policies, labor law….

This transformative capability of embeddings empowers us to break down the silos between seemingly unrelated documents, thereby enhancing our ability to draw meaningful insights, make data-driven decisions, and harness the full potential of document intelligence. In essence, embeddings are the bridge that enables us to unlock a vast quantity of knowledge that was previously hidden by the limitations of traditional data analysis methods.

Personal Data & Social Data

Aggregating data about people - while respecting privacy laws- will also be possible at scale, starting with public social data that we all create on a daily basis.

To illustrate, I asked a question to ChatGPT: 'Given the profiles of people I follow on X/Twitter (for which I provided a few pages of data), what insights can you offer about me?'

Here's what it responded with:

It is clearly not perfect… but I doubt any retailer or recruiter I visit would have such fine grain insights.

Share

The potential applications are vast.

• In retail this translate to enhanced customer intelligence and personalized offering.

• In healthcare, it could mean correlating specific lifestyles with certain diseases, leading to preventive recommendations.

• In human resources, envision leveraging such insights during the recruitment process

Conclusion

Although George Colony’s insights are great and realistic, I think we are at a beginning of new era for data intelligence. The future of AI and the future of business, transformed by AI will come from harnessing the power of documents and texts.

This will enable businesses to gain deeper insights/intelligence and to refine their product and services with this newly uncovered intelligence.

I often compare GPT to a calculator or a spreadsheet for “text and ideas”. The introduction of LLMs will be as revolutionary as the invention of the calculator, reshaping our understanding and use of (unstructured) information.

Subscribe now

Photo by Brett Jordan on Unsplash

Photo by Charlie Wollborg on Unsplash

I worked recently on a project where the primary objective was to construct a recommendation system, and we started testing GPT for multiple choice questions.

The goal was to see whether GPT was able to correctly rank options and to compare these options with what experts would recommend.

Beside the ranking, we also asked for a confidence level.

Very quickly we discovered that the results were influenced by the order in which we provided the options. This is consistent with this great study done on the topic, Lost in the Middle: https://arxiv.org/abs/2307.03172

To gauge the extend of this side effect, I made a rudimentary test, asking GPT to rank Excedrin, Advil and Tylenol as a treatment from migraine and here are the results:

Here are the learnings:

1. The consistent ranking order is Excedrin, Advil, then Tylenol, indicating no "lost in the middle" effect.

2. The confidence rankings appear unreliable, with no discernible bias based on the order of options presented.

3. GPT struggles with non-linguistic tasks, such as statistics, mathematics, planning, and the like.

My hypothesis is that the order does have an influence, but only when the options are semantically very similar.

Given that the client's issue pertained to a niche domain (precision oncology), the options, when viewed from a high-level perspective, were indeed closely related.

After all, for GPT there is nothing semantically closer to a biomarker than another biomarker.

Photo by Charlie Wollborg on Unsplas

It doesn’t work this way.

First, just considering the data size. While GPT's size is impressive, it's estimated to be under 1 million terabytes.

In contrast, the vast expanse of the internet size is more in the billions of terabytes. Comparatively, it's kind of contrasting a water bottle with fifty Olympic-sized swimming pool. The difference is huge.

Going beyond data size, LLMs exhibit an amazing understanding of language, enabling them to abstract and infer based on context. This capability doesn’t come from accessing a vast database of facts but from what we call "parametric knowledge." This enables LLMs to generate the most probable text sequences in response to a given input.

However, it's crucial to distinguish: parametric knowledge isn't synonymous with facts. It is similar to what we use to call “tacit knowledge” in the earlier days of AI.

Think of it this way: an LLM's knowledge reservoir is similar to our human memory. We recall certain facts or phrases from all we've learned, connecting ideas and thoughts seamlessly. Interestingly, common sense facts (like: “fever is bad”) that we experience (and document) over and over is the type of knowledge, intuitive and non implicit, that makes its way into the parameters.

Most of the time, our recollections are accurate, but occasionally, just like anyone, we might mix up details or even embellish a bit. After all, who hasn't done that once in a while? 😉 I certainly have, only to be met with friends or family asking: "Show me the data! You are making up the numbers!".

This is what, in the context of LLMs is called hallucinations and this is this is the #1 reason why, in B2B, LLMs and generative AI haven't (and cannot?) succeeded for tasks heavily dependent on content generation.

Using only parametric knowledge for answering questions is like asking your fresh MBA assistant to help only using its memory , not giving they access to a library or to Google. It is like going back before writing and printing and relying only on “brain to brain” knowledge transfer.

The invention of writing and printing has indeed been a condition for business. Think about it: banking accounts, laws, contracts, records, product descriptions…. there is no business without factual recorded knowledge. Business and B2B demands explicit knowledge.

While it's OK and even desirable to have creative liberties in advertising copy or artistic endeavors (hallucinations are not really a problem for image generation), the stakes are undeniably higher in B2B. B2B applications demand precision, accountability, and trustworthiness and answers must not only be accurate but also verifiable and dependable.

The net is that for B2B, an application cannot depend on parametric knowledge only and has to rely on external sources, using an architecture called Retrieval Augmented Generation.

This architecture works this way:

1- a body of knowledge, specific to the B2B task is built and encoded, sometimes in different formats (semantic vectors, plain text, knowledge graph). (RAG Knowledge set)

2- at inference time, the user query is compared to the RAG knowledge set and the most relevant articles/part of articles are added to the context of the query

3- the augmented prompt is sent to the LLM with instruction to formulate an answer relying on the information present in the context.

A more developed architecture is described in this document from Andreessen Horowitz. It shows (upper right) the need for a Vector database that encodes custom knowledge and is retrieved and passed as context at inference time.

The consequences for Corporations is that B2B generative applications are pivoting towards data and search challenges:

• Data Collection Concerns:

  • Availability and Context: Do we possess the necessary data, and can we pinpoint the relevant context? Should we enrich the data with public, social, private data that are commercially available?

  • Accuracy and Organization: Is our data precise, and systematically arranged?

  • Completeness: Are there gaps in our data repository?

• Data Structuration Dilemmas:

  • Storage Modalities: Should our data be housed within a file system, a document management platform, an SQL database, a semantic vector database through embeddings, or perhaps a knowledge graph?

  • Granularity of Data: What should be the optimal size of our "data chunks"?

  • Embedding Strategies: Which technique is most suitable for embedding our data?

• Search Challenges:

  • Volume: How many documents should our search retrieve? Keep in mind that although LLMs allow larger and larger contexts, most business models are “pay per token” and the larger the context the more we pay. Plus it is not clear that very large context

  • Matching: How can we best align the user's "prompt" with our knowledge repository?

• Opportunities with NLP:

  • At every juncture of this process, there are potential enhancements and optimizations that Natural Language Processing (NLP) can offer.

Overall this is not really new. The challenges and approaches resonate with AI & NLP projects that industry professionals have been grappling with since the 1980s.

Share

I have always been fascinated by the “Library of Babel”, the short story from Borges: https://www.amazon.com/Library-Babel-Jorge-Luis-Borges/dp/9333185879/

and I think it is an interesting book to revisit as we try to understand and make sense of LLMs and ChatGPT.

If you have not read it and want to, please stop reading this post, read the book and come back as I am going to spoil the main story in the next section.

In this novel, the author describes the universe as a vast and virtually infinite library. This library contains every possible 410-page book of a certain format and character set. Essentially, the library holds all possible combinations of 25 symbols: 22 letters, the comma, the period, and the space.

The librarians claims it contains all possible combinations of their symbol set and thus, all knowledge. Due to the infinite number of possible combinations, almost all the books are complete gibberish. The library's inhabitants go on various quests to search for meaningful books, creating theories and cults around the search. Some hope to find a book that indexes all the others, some seek their own personal histories, some even worship the gibberish books.

The narrator concludes that the library has no end. He implies that the library is essentially chaotic, that any order or meaningfulness is imposed arbitrarily by human minds. The story thus explores themes of infinity, information overload, and the human desire for pattern and meaning.

Although there are fundamental differences on how GPT is constructed and the way it works (GPT does not have pre-stored outputs and generates responses dynamically), looking at GPT with a Borges’s prospective helps make sense of its intelligence (and lack of) and is instructive on how to use it and why not to take its responses for the truth. At least it helped me !

If we imagine a library where every possible book that can be written exists, then GPT can be viewed as a mere search into this infinite possibilities. It is not a random search but one using as guidance, traces made by every past conversation, every past story, every documented moment of human life used in its training.

So what a user of GPT does is provide the beginning of a book, a first sentence and lets GPT find the most probable book, using as trace any sentence that has been written before.

Most of the time, ChatGPT responses make sense and even look magical.

But it does produce wrong or made up results from time to time, which people call hallucinations - a better marketing term for failure of error.

Keep in mind, these fondation models have been trained by everything that is on the internet, where one finds news, stories but also fictions and … crap.

When you’re using ChatGPT, you’re embarking on a journey to find the book that answer your question, following the traces of everything that has been written/tried before on the internet.

There are use cases which are fairly safe: As a non English speaker I use it to check my spelling and grammar, sometimes to translate. It is also great for “non additive tasks” such as summarization and text simplification.

Any use case that highly involves “generation” i.e advice/recommendation system, fact checking, specially in domain with a long history of biais ( Medical, Legal, HR …) would require a full solution around GPT including grounding, fine tuning, reinforcement learning, ensemble … there is an entire industry building guardrails.

It works specially well when taking roads that people have taken a lot.

I started this AI blog journey with a literature analogy and would like to close with another master piece of poetry: The road not taken (https://www.poetryfoundation.org/poems/44272/the-road-not-taken)

Share

@dominiq

Subscribe now