• Therapy (more on this one later) is the new top use case. There are two other entrants in the top 5: “Organizing my life” and “Finding purpose.” These three uses reflect efforts toward self-actualization, marking a shift from technical to more emotive applications over the past year.

• “Personal and Professional Support” is now the largest theme by far, stealing most of its new ground from “Technical Assistance & Troubleshooting.”

   
• Therapy/companionship (#1).
• Organizing my life (#2).
• Enhanced learning (#4).
• Healthier living (#10).
• Creating a travel itinerary (#24).
• Disputing a fine (#83).
• More and more of us are using LLMs to find purpose and improve ourselves.
• Therapy involves structured support and guidance to process psychological challenges, while companionship encompasses ongoing social and emotional connection, sometimes with a romantic dimension.
• Many posters talked about how therapy with an AI model was helping them process grief or trauma. Three advantages to AI-based therapy came across clearly: It’s available 24/7, it’s relatively inexpensive (even free to use in some cases), and it comes without the prospect of judgment from another human being.
• A growing number of professional services are now being partially delivered by generative AI—from therapy and medical advice to legal counsel, tax guidance, and software development.
• The highest new entry in the use cases was “Organizing my life” at #2. These uses were mostly about people using the models to be more aware of their intentions (such as daily habits, New Year’s resolutions, and introspective insights) and find small, easy ways of getting started with them.
• The other big new entry is “Finding purpose” in third place. Determining and defining one’s values, getting past roadblocks, and taking steps to self-develop
• There was much disagreement and discord about whether the LLMs are a boon or bane for our own ability to think.
• This fear was manifested more explicitly on behalf of young people. Many people expressed concerns about the impact on college education (#23) in a world in which essays can be on-point, high-quality, and on-demand.
• The hottest cynical take is about the perceived political correctness of the LLMs. This may have been more pronounced over the course of the 2024 election year in the U.S. Gen AI lost at least this user as a result: “I cancelled my subscription for this reason. I do not want to support a tool that tries to turn the world more timid and prim.”
• Enthusiasm for gen AI was frequently met with warnings about what Big Tech would do with the data it’s harvesting
• another common complaint was that the LLMs don’t know enough about their users,
• More seem to understand the fundamental point that for AI to be useful to humans, we need to be clear about our intentions: “Of course it’s only helpful when I already know what I want to do.”
• . One common and more nuanced forecast was the desire to see LLMs move from advice and information to doing, i.e., agentic behavior

20 more annotations...

• obsession misses a lot of truths about how culture is actually created and what steers it to be “good” or “bad.”
• workplace culture represents how we work together
• culture emerges through complex interactions among leadership behavior, team dynamics, and individual experience. While executives set the tone through their actions, culture is ultimately cocreated through daily interactions at every level.
• 1. Focus on nurturing dependability-based trust. Trust forms the foundation of an effective workplace culture. Dependability-based trust — the confidence that colleagues will follow through on commitments — is the base that effective teams are built on.
• 2. Cultivate a first team mindset. One of the most transformative cultural shifts happens when leaders embrace the concept of the first team. This approach, popularized by author and consultant Patrick Lencioni, involves executives prioritizing their peer team over their functional silos — silos where, as Lencioni puts it, they’re “the papa bear or mama bear” to employees they already feel deep loyalty to.
• 3. Leverage personal user manuals to build connection. Personal user manuals have emerged as a powerful tool for accelerating trust and understanding, especially in distributed teams. These brief self-disclosures help colleagues understand how each person works best. Documents can include reflective notes on a person’s communication preferences, their values and what motivates them, strengths and growth areas, and personal context that might impact work
• 4. Establish team agreements. Effective workplace cultures don’t rely on rigid policies but rather on clear agreements within each team of collaborators about how work gets done.
• 5. Focus on outcomes, not attendance. Perhaps the most fundamental cultural shift needed in today’s workplace is moving from monitoring activity to measuring impact. Being driven by hustle culture perpetuates the tired belief that showing up matters more than delivering results.
• 6. Design intentional gatherings that build belonging. While daily colocation isn’t necessary for a strong culture, intentional in-person gatherings play a vital role in building connection and a shared identity. At Future Forum, the pandemic taught us the vital importance of quarterly gatherings designed with clear purpose.

7 more annotations...

• La nouvelle entité portée par Adecco et Salesforce vise précisément à offrir un cadre structurant à cette transformation, en proposant aux organisations une approche industrialisée de l’intégration des agents IA dans les effectifs.
• L’offre se présente comme un service intégré aux outils métiers que les entreprises utilisent déjà, notamment Slack ou les produits CRM de Salesforce.
• Elle a pour but de permettre aux décideurs de mieux piloter l’allocation des ressources, en apportant une visibilité sur la répartition des tâches entre humains et IA, la possibilité de simuler des scénarios de transformatio
• Avec une baisse de chiffre d’affaires de 3 % en 2023, à 22,7 milliards d’euros, le groupe cherche à se repositionner sur des segments porteurs.
• Pour Salesforce, ce partenariat s’inscrit dans sa stratégie d’évangélisation de l’intelligence artificielle dans les usages quotidiens de ses clients.
• Cette première « agence d’intérim algorithmique » ouvre ainsi la voie à une nouvelle manière de penser l’organisation du travail, fondée sur la complémentarité entre les compétences humaines et les capacités des systèmes intelligents.
• elles seront les modalités de gouvernance de ces agents ? Comment garantir leur bon usage, leur transparence et leur éthique ? Quels effets à moyen terme sur l’emploi traditionnel ?
• urd’hui. La frontière entre collaborateurs humains et assistants virtuels s’estompe progressivement. Et avec cette initiative, Adecco et Salesforce entendent bien organiser cette convergence, en industrialisant le recrutement,

6 more annotations...

• They’ve accepted that agility isn’t a one-time move. It’s a muscle, and they’re shifting their focus from “reacting fast” to “designing smart.”
• This approach of creating intentional space between reaction and redesign is becoming standard operating procedure for HR teams

• The teams that thrive aren’t just reacting faster. They’re designing systems that anticipate the next pivot before it arrives.

   
• Well, those days are a distant memory as we’re now operating in what one client calls “a perpetual beta.” The org chart you finalized last quarter? Already outdated. That three-year strategic plan? More like a wish list with a distant north star.
• As one CHRO bluntly stated, “We spent decades building organizations designed for efficiency and predictability. Now, we need organizations built for resilience and reinvention.”
• 66% of American employees are experiencing some sort of burnout in 2025. The data reveals that younger generations are facing significantly higher rates of burnout, with 81% of 18- to 24-year-olds and 83% of 25- to 34-year-olds reporting burnout, compared to just 49% of those aged 55 and older

• 1. They’ve redefined stability.

   

  Instead of promising employees that nothing will change, they’re creating psychological safety around how things change.

• They’re building learning ecosystems.

   

  As the half-life of skills continues to shrink, progressive companies are creating continuous learning environments where knowledge flows in all directions.

• They’re making transparency a competitive advantage.

   

  In a world of constant change, information asymmetry breeds mistrust.

• They’re reimagining career development.

   

  The linear career ladder is giving way to “career constellations,” flexible pathways that allow for lateral moves, skill expansion and even temporary role experiments.

• Scenario planning as a team sport

   

  Rather than relegating scenario work to the strategy team, forward-thinking organizations are involving broader groups in imagining different futures and identifying early warning signs.

• Seeing if your pulse is racing

   

  Annual engagement surveys can’t capture the rapid shifts in employee sentiment. Real-time feedback tools are becoming essential for sensing when pivot fatigue is reaching critical levels

• Capability mapping for total workforce management

   

  Instead of focusing solely on current roles, organizations are cataloging transferable skills and creating visibility into how these skills can be deployed across different business needs
• We’re in a world where strategies expire fast, expectations are fluid and transformation isn’t a destination; it’s a working draft.
• The real opportunity in 2025? Stop resisting the pivot and start owning it

13 more annotations...

• “On the one hand, Musk selling X to xAI looks like him transferring money from one of his pockets to another,” said Rik Turner, senior principal cybersecurity analyst at Omdia. “However, by having the still highly influential social media site to his AI firm, we must ask whether he intends to use all of X’s content as training data for xAI.” 
• Data from the social media platform has already been used by xAI to train its AI models and chatbot Grok,
• "This development definitely raises data privacy concerns,”
• "Training AI models on X content introduces another major issue: the reliability of the data,” he added. “Unlike vetted sources, X has no systematic fact-checking, meaning AI systems could absorb and amplify misinformation, biased narratives, or even harmful content. Without safeguards, this could lead to AI models generating misleading or inaccurate information, reinforcing echo chambers, and eroding trust in AI-driven insights." 
• Even when a social media platform is sold to an AI company and it raises a lot of questions and concerns, the movement goes in parallel as other large corporations integrate AI across the board such as Google and Microsoft.
• “We should consider to what extent the xAI will have access and utilization of X’s data,”
• since AI needs to access data to do its job, it’s up to businesses to employ guardrails to ensure sensitive information isn’t being used. 
• “By acquiring X – xAI gains access to this treasure trove of personal data, as the social media platform has over 600 million active users, and much of this data is likely highly detailed, including not just user posts but also behavioral data,
• “This is a prime example of how the lines between user privacy and the use of personal data for AI research and development is becoming more blurred.
• “The option to opt out by adjusting settings or making an account private shifts the burden onto individuals rather than ensuring true informed consent,”

8 more annotations...

• Dans cette course mondiale, la France a pris position. Elle a investi, structuré, accompagné. Elle a semé. Mais reste à savoir si elle saura récolter à grande échelle.
• la France dispose de tous les ingrédients scientifiques et entrepreneuriaux pour devenir un acteur de premier plan de la deeptech mondiale, mais elle ne franchira ce cap qu’à condition d’opérer un changement d’échelle rapide et structuré. 
• Le risque est élevé, l’incertitude forte, mais la valeur ajoutée — économique comme sociétale — est considérable. À ce titre, la deeptech dépasse le cadre de l’innovation classique pour devenir un outil de souveraineté industrielle et stratégique.
• L’écosystème s’est structuré avec plus de 2 500 startups deeptech aujourd’hui actives sur le territoire, contre un peu plus d’un millier il y a cinq ans. En 2023, les montants levés ont atteint 4,1 milliards d’euros, contre 1,5 milliard en 2019.
• l’État a investi 4,5 milliards d’euros sur la période, combinant subventions, concours, prêts, fonds d’investissement et accompagnement stratégique
• la France se heurte à un mur : celui du passage à l’échelle industrielle. À partir des séries B et C, le capital disponible se raréfie.
• Les sorties, cessions, introductions en bourse, sont peu nombreuses, ce qui freine les flux de capitaux.
• En 2023, les États-Unis ont investi 46 milliards d’euros dans les technologies deeptech, la Chine 17, contre 4,1 pour la France.
• Il ne s’agit plus seulement de créer des startups, mais de bâtir des champions industriels capables de rivaliser à l’échelle mondiale.
• Le transfert de la recherche vers l’industrie doit être accéléré, les interfaces entre laboratoires et marchés mieux structurées, les verrous technologiques levés plus rapidement.
• Trop d’innovations restent confinées au stade du prototype. Pour qu’elles deviennent des produits transformant les filières industrielles, il faut des usines, des équipements, des chaînes de production.
• Le rôle des grands groupes est également central : non plus en simples partenaires de recherche, mais comme investisseurs, acheteurs, et vecteurs de diffusion technologique.
• En définitive, la deeptech n’est pas seulement un enjeu de startup nation. Elle est une affaire de souveraineté, de résilience industrielle, de projection géopolitique.

11 more annotations...

• First of all, AI’s potential as a GPT lies in its versatility. GPTs are broadly applicable across sectors, improve rapidly, spur complementary innovations and fundamentally reduce the cost of critical economic activities. Unlike specialized technologies, AI can tackle a vast array of tasks—from summarising emails to designing board games, from writing code to finding new molecules.
• . Downstream, we are seeing huge innovations in hardware; performance per dollar of Nvidia GPUs has been improving by around 30% per year.
• Upstream we’re seeing wild experimentation. In AI-based middleware we have LangGraph and Wordware both helping developers create sophisticated AI agents. Apps like Cursor help coders work faster, Granola simplifies meeting notes and Elicit speeds up literature reviews for researchers.
• AI promises to lower the cost of intelligence — a core input in modern economies. When agents can do a task, they do so at one-thirtieth of the cost of a human.
• For AI to succeed as a GPT, it must drive a noticeable increase in productivity growth.
• History shows that this can take time and follows an S-curved trajectory. In the short term, the impacts of a GPT are marginal across an economy, although individual firms can do well. In the medium term, as the technology becomes more widespread, firms are able to retool around the technology and a plethora of upstream and downstream businesses emerge. This all drives growth. In the long term, a technology’s contribution to productivity growth flatlines as it becomes normalised.
• In other words, it’s the medium term which matters, the point at which a technology is far enough dispersed into firms in an economy to have a systematic impact on productivity.
• Economists are rather split about what AI might mean to us economically. There’s a wild range of estimates looking at labour productivity growth alone.
• Goldman Sachs, for instance, estimates a 1.5% increase in labour productivity over 10 years for 2023. To make sense of this, let’s turn it into real dollars. US GDP was $27.7 trillion in 2023. This could mean that by the end of the tenth year, US GDP would be between $6 and 7 trillion dollars higher than it otherwise would have been, yielding a cumulative $25-30 trillion dollar increase over the decade.1
• t, I’ll define “disappointment” as AI failing to boost labour productivity by 2% by the year five, starting the clock in January 2023. I’m choosing 2% as it’s above most estimates,
• My main line of thinking is that productivity impacts follow the deployment of a technology across the economy and AI can get deployed much faster than the railway: we’re richer and much of the underlying infrastructure, internet and computers, is in place. In addition, the organizational dimensions of the modern economy favor fast adoption
• This might suggest firms have the capacity and managerial expertise to move quickly if they need to. In addition, capital markets could be more punishing – providing a negative incentive – to slow movers.
• , it would mean adding a cumulative $2.5-$2.7 trillion of US GDP between 2023 and 2028, with US GDP about a trillion and change higher in 2028 than it might otherwise have been.
• In the scenario, I am going to outline below, the productivity impact of AI could be limited by several key challenges, even as AI companies themselves grow rapidly. The deployment of AI across the economy might not translate into the broad-based gains we expect, at least not within the five-year timeframe starting January 202
• in the U.S., large firms with over 500 employees drive 70% of GDP. If AI’s reach stays confined to smaller entities, it won’t lift aggregate productivity significantly.
• Conversely, large firms—the economy’s backbone—often lag in fully embracing AI.
• Instead of overhauling operations, they opt for modest gains, like automating customer support or refining HR processes.
• These dynamics—uneven adoption, large-firm caution, and market jitters—might prevent AI from boosting labor productivity by 2% by 2028. This could leave U.S. GDP $2.5–$2.7 trillion below projections for a true general-purpose technology.
• AI products like ChatGPT and Cursor have seen impressive uptake—ChatGPT reaches 400 million weekly users and Cursor is generating $100 million in revenue within two years.
• These successes – and they are huge – are successes for the entrepreneurs building these business. But it is still small beer against the mighty $29 trillion US economy.
• Thirty years into the internet boom, the digital economy contributes 10% to US GDP. Expectations for AI are much, much higher.
• Anthropic and OpenAI forecast collectively $65 billion in revenue by 2027. Assume they grow 30% to 2028. This would take revenues to $80 billion, of which perhaps two-thirds, so roughly $50 billion and change.
• David Cahn at Sequoia Capital reckoned that Ai firms would need to find $600 billion on revenues to justify 2024’s investments in data centers.
• Even if AI becomes inexpensive, the economy may struggle to absorb vast amounts of “intelligence.” We’re already seeing this happen. Large companies, who make up the bulk of the real economy, are starting their genAI implementations to eke out efficiencies from low-value uses, like customer support.
• more and better machine intelligence will be available increasingly cheaply. Look at the price drops we’ve seen in GPT-4 tokens.
• But companies and industries need to absorb this intelligence. And the economy might have natural saturation points that limit uptake. It took companies, and then industry, two decades to change their processes to make use of electricity in the factories.
• If AI is as powerful as it might be, integration into firms could be more complex, not simpler.
• This isn’t to say the individuals in the large firms, which make up the bulk of the US economy, wouldn’t start to make smart use of AI. There is already evidence they are. Indeed, a recent survey of surveys by the Fed suggested that worker-level use of AI was higher than firm-level use.
• AI tools are being used as augmentations to workers, training them to deal with the jagged frontier might delay roll-outs. For developing highly automated systems, higher reliability requirements might restrict activities to well within that frontier.
• How many firms would really benefit today from a thousand PhD chemists showing up at their door?
• Only industries that are already heavy in R&D, like pharmaceuticals, might salivate at the prospect but most industries don’t really know what to do with the capabilities.

• The AI boom is driven by massive capital expenditures. Amazon, Microsoft, Google and Meta are projected to drop $320 billion on capex this year, with AI chip demand expected to grow from $114 billion in 2024 to $280 billion by 2028, according to Morgan Stanley.

  Morgan Stanley roughly estimates $520 billion by 2028. Assuming a 50% margin, it implies $1 trillion in new revenues that need to be found in that year, well above my threshold for disappointment.
• Alongside Big Tech’s balance sheets, hundreds of billions of dollars of financing will come from debt markets to buy the land, build the buildings and finance the power generation these data centers need.
• When the debt-fuelled telecoms bubble burst in the early 2000s, failing firms left behind a vast network of “dark fiber” cables, ducting routes and physical infrastructure that became a critical resource for subsequent internet growth. The OECD noted that while painful, this infrastructure legacy became the “necessary substrate” that enabled streaming, cloud computing and modern AI systems that now consume 85%+ of the original bubble-era capacity.
• Unlike the telecom bubble, where unused fiber ducts retained value, high-end GPUs can depreciate quickly – between 2-5 years depending on usage. Those GPUs and data centers may not be useful if there is a bust.
• The numbers are now getting so large from just even a couple years ago that the returns on invested capital are really now beginning to turn down pretty hard for these companies.
• With 65% to 80% of data center development funding reliant on debt, a wave of defaults or write-downs in this debt could deepen the crisis, as jittery investors pull back from riskier assets.
• Historically, markets take several years, not less than five, to recover from a real asset bubble
• But venture class risk has turned into broad stock market risk, because 30+% of the US stock market now rides on tech firms astride the genAI wave
• In this world, AI’s productivity impact could lag due to slow adoption, limited demand elasticity and over-investment risks.
• For example, what if the US can’t get enough power infrastructure on stream to meet enterprise demand for genAI? I touched on this risk in an op-ed in the New York Times last year. That would likely result in price rises or supply constraints that could slow down getting results from AI.
• Today’s essay outlines a plausible, low-probability scenario for AI having slower than expected effects on the economy resulting in a cascading disappointment.
• It is my baseline view that AI, on its current iterative trajectory, will move faster into the economy than previous technology waves.2

42 more annotations...

• La dureté de patrons comme Musk, ou avant lui Steve Jobs, est légendaire. Ils sont impitoyables. Ils exigent l’impossible. Ils veulent toujours plus. Ils critiquent très durement le travail de leurs équipes. Ils sont cassants, arrogants, colériques, jamais satisfaits.
• Rien ne permet en effet de dire que le management de Musk et Jobs, et plus généralement celui des groupes ultra-exigeants, est inhumain.
• Cette association est généralement la traduction d’une hostilité à la performance, que l’on conçoit nécessairement comme toxique.
• on peut évoluer dans un environnement extrêmement exigeant, ultra-compétitif, sans pour autant qu’il soit inhumain.
• La dureté d’un manager est la traduction d’une exigence, et donc de respect.
• Pour un Jobs, cette exigence n’a de sens que parce que la personne qu’il critique – ou qu’il attaque, plus exactement – est quelqu’un de talentueux, et qui le sait.
• Jobs et Musk ne travaillent pas avec des médiocres, seulement avec des gens exceptionnels; donc ils estiment ne pas avoir à s’embarrasser de fioritures.
• Le travail avec Musk sera un enfer, mais Carlson jouera un rôle essentiel dans le lancement de la fusée Falcon 9, véritable rupture dans le transport spatial.
• Malgré la dureté de l’expérience qu’ils imposent à leurs équipes, Jobs et Musk ont su susciter une très grande loyauté de la part de nombreux cadres très talentueux, ce qui suggère que leur leadership a eu un impact positif et motivant.
• A contrario, le management sans exigence est inhumain
• ? Ce sont dans les organisations médiocres que les gens sont malheureux. Un chef qui n’exige rien de ses équipes est inhumain. Il accepte la médiocrité, il n’est pas respecté; mais l’inverse est aussi vra
• Des environnements comme Apple ou SpaceX ne sont pas faits pour tout le monde, c’est évident, parce que tout le monde n’a pas les mêmes besoins, les mêmes envies ni les mêmes aspirations. Ils sont pourtant profondément humains car rien n’est plus humain que de se consacrer corps et âme à la résolution de défis majeurs.

10 more annotations...

• This tech is also poised to change how workers are managed. AI systems and agents are increasingly able to oversee workflow, and how employees and AI work together on those tasks.
• Yet while the percentage of organizations adopting AI has jumped from 55% in 2023 to 78% in 2024, according to McKinsey, employees nevertheless display significantly less enthusiasm about using AI in their daily activities.
• trust is in short supply
• 52% reported feeling “worried” about the impact of AI in the workplace and 33% said they felt “overwhelmed,” compared to 36% who described themselves as “hopeful.”
• Edelman found that 35% of respondents said they rejected AI compared to 30% that embraced it.
• According to another Pew Research survey, 71% of Americans opposed companies using AI to make final hiring decisions, and 61% opposed using it to track workers’ movements.
• roughly half believed their job might disappear in the next decade.
• According to PwC’s recent Responsible AI survey, only 58% of companies have completed a preliminary assessment of AI risks at their organization.
• they see these tools as lacking benevolence in how they make recommendations and decisions, and worry about whether AI will “care” and consider employee interests.
• it represents the fundamental characteristic of the trustee (AI in this case) to do good for the trustor (employees in this case) as basis for trust to emerge.
• a signal to organizational leaders that confidence in them might be lacking, too.
• sentiment around AI also reflects how workers feel about leadership: how they think leaders will use these tools and their faith in leadership’s motivations in deploying AI.
• they need to ask how AI will shape how employees perceive benevolence.
• AI systems are increasingly adept at recognizing patterns in human interactions. This capability is what makes them so effective to function (semi-)autonomously in areas like customer service, personalized recommendations, and workflow management. By analyzing vast datasets of human behavior—encompassing text, voice, and even gesture data—AI systems can make nuanced inferences about human preferences, communication styles, and responses, and subsequently make decisions.
• Because of this, these systems can also communicate in ways that can seem human-like—they can adapt to emotional cues in conversations and present as empathic, stern, or encouraging, depending on their instructions.
• Because AI mirrors back what it sees in the data and takes cues from whoever is giving it instructions, it can reflect both the lived and stated values of a company.
• Organizations can train AI that provide recommendations to appear benevolent—or not—through the use of targeted prompts and requests. An AI might be instructed, for instance, to identify and prioritize the interests of the stakeholders that are involved in a business decision.
• Even so, employees aren’t likely to experience these interventions as evidence of AI’s benevolence—nor should they—because they generally think of that as a human quality and one out of reach of these tools.
• AI systems aren’t—and can’t be—benevolent. AI does not have the ability to be willing to engage in a positive orientation to serve the interests of an interdependent other. Indeed, these systems operate through algorithms that follow statistical patterns and predefined goals without any self-driven purpose; and as such there is no intent in the internal logic of the AI engine.
• 1) humans are better than chance in detecting whether emotional cues are real, and 2) research reveals that benevolent trust is hard to build if the intentions of the other party are in doubt.
• . Ensuring that AI is accepted by employees requires that they are able to understand it and how it works
• First, to build an environment where employees can confidently adopt AI, organizations must promote AI’s cognitive trust factors—reliability, accuracy, and transparency.
• Second, leaders need to demonstrate genuine concern for employee well-being and fairness, because AI can’t. Ultimately, leadership will be deciding how AI is used and empowered, and where it isn’t.
• Third, the optimal solution for the lack of trust in AI requires an “AI-leader combination” that ensures that employees feel not only technically confident in AI but also reassured that it will serve their interests in respectful and fair ways.
• Creating a work context where AI is both reliable and emotionally trusted requires a deliberate alignment of technological interventions to enhance cognitive trust with benevolent and transparent leadership.

23 more annotations...

• I didn’t really care about being a founder, I just wanted an operational job of some sort,
• I’d always been really interested in AI, I worked in the AI lab when I was an undergrad and nothing was working at the time, but I always thought it was the coolest thing, and obviously the structure and plans of OpenAI have been some good and some bad, but the overall mission of we want to build beneficial AGI and widely distributed, that has remained completely consistent and I still think is this incredibly important mission.
• So we had no business model or idea of a business model and then in fact at that point we didn’t even really appreciate how expensive the computers were going to need to get, and so we thought maybe we could just be a research lab
• not just the research problems, but the like, “What should our structure be?”, “How are we going to make money?”, “How much money do we need?”, “How do we build compute?”,
• Like the GPT-2 release, there were some people who were just very concerned about, you know, probably the model was totally safe, but we didn’t know we wanted to get — we did have this new and powerful thing, we wanted society to come along with us.
• Now in retrospect, I totally regret some of the language we used and I get why people are like, “Ah man, this was like hype and fear-mongering and whatever”, it was truly not the intention. The people who made those decisions had I think great intentions at the time, but I can see now how it got misconstrued.
• Because we thought we were going to be a research lab. We literally had no idea we were ever going to become a company. Like the plan was to put out research papers. But there was no product, there was no plan for a product, there was no revenue, there was no business model, there were no plan for those things.
• Of course we would have set it up differently, but we didn’t know everything we knew now and I think the price of being on the forefront of innovation is you make a lot of dumb mistakes because you’re so deep in the fog of war.
• , we launched GPT-3 in the API, and at that point we knew we needed to be a company and offer products, and GPT-3 was doing fine. It was not doing great, but it was doing fine. It had found real product-market fit in a single category. This sounds ridiculous to say now, but it was a copywriting, was the one thing where people were able to like build a real business using the GPT-3 API.
• Then with 3.5 and 4, a lot of people could use it for a lot of things, but this was like back in the kind of dark ages and other than that, the thing people were doing with it was developers. We have this thing called The Playground where you could test things on the model, and developers were trying to chat with the model and they just found it interesting.
• we thought we could tune the model to be better and then we had this thing that no one knew about at the time, which was GPT-4, and GPT-4 was really good. In a world where the external world was used to GPT-3 or maybe 3.5 was already out, I don’t remember, I think it was by the time we launched ChatGPT — it was, but not 4. We’re like, “This is going to be really good”.
• we finished training GPT-4 in, I think August of that year, we knew we had this marvel on our hands, and the original plan had been to launch ChatGPT, well, launch a chatbot product with GPT-4,
• A lot of people thought it was going to totally flop, I thought it was going to do well, no one obviously thought it was going to do as well as it did, and then it was just this crazy thing. I think it’s amazing.
• was already a pretty viral moment, and if we had launched ChatGPT with 4, it would have been a mega, mega viral moment. One of the craziest things that I’ve ever seen happen in Silicon Valley was that six months where we went from basically not a company to a whole big company
• We had to build out the corporate infrastructure, all the people and the stuff to serve this, it was a crazy time, and the fact that there was a little bit less demand because the model was much worse than 4 turned out to be, that was really important.
• now we have these things that are replacing Google for a lot of people connected to the Internet,
• But, what I wanted was to get to run an AGI research lab and figure out how to make AGI. I did not think I was signing up to have to run a big consumer Internet company

• At this scale, yes, it is maybe the biggest. I guess we are the biggest company since Facebook to be started probably.

      

• I don’t think we’ll be a platform in a way that an operating system is a platform. But I think in the same way that Google is not really a platform, but people use sign in with Google and people take their Google stuff around the web and that’s part of the Google experience, I think we’ll be a platform in that way.

      

• when you talk about serving billions of users and being a consumer tech company. This means advertising. Do you disagree?

      

  SA: I hope not. I’m not opposed. If there is a good reason to do it, I’m not dogmatic about this. But we have a great business selling subscriptions.

      
• I am more excited to figure out how we can charge people a lot of money for a really great automated software engineer or other kind of agent than I am making some number of dimes with an advertising based model.
• I’d be much more excited to try than traditional ads is a lot of people use Deep Research for e-commerce,
• And maybe there’s a tasteful way we can do ads, but I don’t know. I kind of just don’t like ads that much.
• I’m not going to say what we will and will never do because I don’t know, but I think there’s a lot of interesting ways that are higher on our list of monetization strategies than ads right now.
• But it was a lesson for us about when we leave a feature hidden, we left chains of thought hidden, we had good reasons for doing it, but it does mean we leave space for somebody else to have a viral moment. And I think in that way it was a good wake-up call.
• : I still think there can be big risks in the future. I think it’s fair that we were too conservative in the past. I also think it’s fair to say that we were conservative, but a principle of being a little bit conservative when you don’t know is not a terrible thing.
• . If you want something deterministic, you should use a database.
• And I think the problem, the interesting problem is how do you get models to hallucinate only when it benefits the user?
• First of all, most of the regulation that we’ve ever called for has been just say on the very frontier models, whatever is the leading edge in the world, have some standard of safety testing for those models. Now, I think that’s good policy, but I sort of increasingly think the world, most of the world does not think that’s good policy, and I’m worried about regulatory capture
• But for sure, you don’t want to go put regulatory burden on the entire tech industry. Like we were calling for something that would have hit us and Google and a tiny number of other people. And again, I don’t think the world’s going to go that way and we’ll play on the field in front of us. But yes, I think saying that fair use is fair use and that states are not going to have this crazy complex set of differing regulations, those would be very, very good.
• I think it’s actually a great thing about humanity that people adapt and just want more and better and faster and cheaper. So I think we have overdelivered and people have just updated.
• The more general one is I think people can cultivate resilience and adaptability and also things like figuring out what other people want and how to be useful for people. And I would go practice that. Like whatever you study, the details of it maybe don’t matter that much.

30 more annotations...