• Understanding the Difference Between AI and Agentic AI
• AI in its current form supports pattern recognition, forecasting, and classification across large datasets. It enhances visibility and sharpens decision making within defined parameters. Agentic AI extends this capability by introducing systems that can plan, coordinate, and execute across multiple steps with limited intervention. The distinction matters. One improves decisions. The other begins to carry them out.
• Where COOs Can Realize Immediate Operational Gains
• For COOs, the practical question centers on where these capabilities intersect with operational priorities. Cost discipline, service levels, throughput, and resilience remain the core measures. AI contributes when it strengthens these outcomes in a sustained way.
• Closing the Gap Between Insight and Execution
• Agentic systems can then act on those forecasts by adjusting procurement schedules, reallocating inventory, or coordinating with suppliers.
• Reframing Customer Operations with Intelligent Systems
• AI can interpret intent, segment requests, and recommend responses. Agentic systems can take the next step by resolving issues, routing cases, and initiating follow up actions across systems
• Building the Right Data Foundation
• . Inconsistent definitions, fragmented systems, and incomplete records limit both AI and agentic performance.

• Clarifying Processes Before Scaling Technology

   
• Operations that rely on informal judgment or undocumented steps are difficult to translate into systems that can act independently.

• Designing Technology for Coordination, Not Complexity

   
• Establishing Governance That Enables Scale
• As systems take on a more active role in execution, organizations must determine where human oversight is required and how decisions are reviewed.
• Preparing the Workforce for a New Operating Model
• Employees spend less time on coordination and more time on judgment, oversight, and continuous improvement
• Moving Beyond Efficiency to Operational Coherence
• The combination of AI and agentic systems allows for a more cohesive operating model, where planning, execution, and monitoring are more closely aligned.
• Scaling with Discipline and Measurable Outcomes
• Initial efforts should focus on areas where complexity and coordination challenges are well understood.

• Embedding AI into the Operating Model

   
• Systems will not be viewed as separate tools, but as integral components of how work is performed.

• A Continuous Effort, Not a One-Time Initiative

   
• AI is not a single initiative. It is an ongoing effort to align data, processes, systems, and people around a more coordinated form of execution.

23 more annotations...

• The term “tokenmaxxing” has been gaining steam in recent months, as tech pros compete to spend the most “tokens,” or data units used by AI to process natural language.
• employee perceptions about the link between AI use and performance assessment are changing, all while standard expectations for AI use at work are still forming.
• For instance, an employee at Meta created an internal dashboard to track AI token spend—revealing employees used 60 trillion tokens in one month
• Barney Hussey-Yeo, founder and CEO of financial assistant app Cleo, acknowledged that “everyone at Cleo is tokenmaxxing” and that he was strategizing for how to encourage even more uptake.
• The practice assumes greater token consumption equals greater business outcomes.
• “For many organizations, this metric of consumption does not distinguish between AI usage and the actual value derived from AI,” Rowan says.
• Rowan suggests tokenmaxxing is at risk of becoming a “vanity metric,”
• the numbers may not ultimately connect to real value for stakeholders and the organization.
• Salesforce is taking a slightly different approach to measuring the value of AI usage, using agentic work units, which measure output and impacts, opposed to token spend.
• For instance, Google and Meta are both building AI usage metrics into performance reviews for engineers.
• How can leaders encourage and then measure AI adoption that drives real business outcomes, while avoiding employees using the tech in a performative way
• usage rates are just a “sliver of the story” that HR needs to be thinking about, alongside employee trust, confidence and readiness.
• Such metrics can tell HR “whether or not adoption is real or it’s just compliance theater,”
• about 16% of employees admitted to pretending to use AI. Nearly a quarter say they feel pressured to use AI in situations they’re unsure about.
• half of workers surveyed said they “have no idea” how AI is supposed to improve their productivity,
• Employees need hands-on opportunities to try the tech and see its influence on their work, so they’re not just comfortable using it because of a company requirement, but rather they are truly bought into its value.
• “If you want your workforce to embrace AI,” Jones says. “you’ve got to stop training them and start enabling them.”

15 more annotations...

• Chapitre premier. Optique et perspective : Ptolémée, Alhazen, Alberti1
• Nous voudrions montrer que l’une des conditions théoriques préalables à l’apparition de la perspective classique, au début du XV^e siècle en Italie, a été la révolution (le mot ne nous paraît pas trop fort) opérée en optique au début du XI^e siècle à Bagdad par Ibn al-Haytham (dont l’occident médiéval a latinisé le nom en Alhacen ou Alhazen), quand ce dernier montra que la vision se produit grâce à des rayons lumineux qui pénètrent dans l’œil et non grâce à des rayons visuels qui en sortent, comme le postulait l’optique géométrique de l’antiquité.
• La thèse peut paraître d’autant plus surprenante qu’Alberti lui-même affirme expressément dans le De Pictura que pour ce dont il traite, il est sans intérêt de choisir entre les deux théories optiques concurrentes, émission par l’œil de rayons visuels comme le voulaient Euclide et Ptolémée, ou intromission dans l’œil de rayons lumineux comme le pensaient Alhazen (nom sous lequel fut traduit en Occident Ibn al-Haytham), ainsi que Vitellion et les autres auteurs européens qui très vite adoptèrent ses idées. Alberti se contente de rappeler les deux possibilités retenues de très longue date par les spécialistes et il s’en tient ensuite au langage traditionnel de la pyramide visuelle issue de l’œil, plus commode pour son propos, car mieux apte à faire comprendre aux peintres les conséquences techniques des données optiques qu’il expose2.
• e Traité d’optique d’Alhazen avait été traduit en latin à la fin du XII^e ou au début du XIII^e siècle3 et soit par cette traduction, soit par la reprise élargie qu’en avait donnée Vitellion, non seulement sa théorie d’une entrée dans l’œil de rayons lumineux s’était bien diffusée dès la fin du XIII^e siècle, mais, quelle que soit l’hypothèse optique que l’on admettait (émission ou intromission), on considérait avec lui que la perception visuelle, en particulier celle des distances et des grandeurs, ne pouvait résulter que d’un jugement complexe comportant même une part de raisonnement. Sa psychologie de la perception visuelle alimentait donc les discussions sur la vision et les différentes variantes des théories de la connaissance scolastiques4. Or c’est cela qui compte pour le développement d’une perspective systématisée, et nous allons tenter de montrer pourquoi.
• Toutefois, malgré cette géométrisation du visuel, il ne faut pas chercher dans Euclide un manuel de perspective. Bien qu’il ait eu la possibilité de réfléchir sur les décors de théâtre, il ne fait expressément allusion dans son optique à aucune pratique artistique

3 more annotations...

• The genius of LLMs is that, in many cases, they seem impressive if you don’t have the expertise and are a pain in the wrong place if you’re skilful
• As of late 2025, LLMs were able to complete human-level tasks only at a 65% success rate, and the study predicts that by 2029, they’ll be able to complete text-related tasks at a success rate of 80-95% at “minimally sufficient quality.” 
• That’s a C-minus score at school; the kind of work that gets you a “meets expectations with support” appraisal at work
• What I am counting on is that high-end knowledge work isn’t going anywhere. Judgement, skill, taste, and metacognition will matter even more than ever before.
• The hard thing about hard things is never the hard thing.
• Reading isn’t the same as gleaning a book summary.

• You can decompose each of these hard things into four constituent parts.

  1. The starting point or idea.

  2. The process of problem solving.

  3. The output or artefact from the process.

  4. The outcome that the process and output achieve.
• he cognitive struggle is an integral part of the creative process
• The cognitive struggle is as much in service of the output as it is an exercise for our brains. It’s what keeps us sharp.
• The biggest fallacy about knowledge work, then, is treating the output as the end-all.
• Design is about solving user problems. Presentations are about telling a compelling story. School is about learning. The point of school surely isn’t to feed a homework goblin.
• Using LLMs to escape the cognitive struggle erodes our cognitive capacity.
• The Wall Street Journal and Harvard Business Review have both reported that LLMs don’t reduce work – they intensify it.
• If your company provides access to platforms like Glean, UX Pilot, Replit, OpenClaw or even subscriptions to Perplexity or Claude, producing outputs is the easiest part. Dealing with them is the hard bit. 
• It’s about dealing with a higher level of entropy in the system – change management, release management, decision-making, communication, and coordination – everything becomes more intense and more brittle.
• the WSJ piece found the use of email, messaging, and chat apps has doubled, while the time for focused work to solve complex problems has fallen by 9%.
• Individual tasks happen faster – you can keep reaching for that generate button – but when did speed and productivity become the same thing?
• if you see your job as merely producing outputs, you’ll be less valuable than someone who owns the process, outputs, and, most importantly, the outcome. 
• If I’m good at something, I’m not delegating that work to LLMs. I’ve honed my craft over several years to the extent that I’m several times better than the basic outputs an LLM produces.
• To make time for the things I’m good at, I’ll use LLMs in the most cynical way possible. I’ll delegate all the boring, box-ticking, Dilbertesque work to generative AI tools
• I’m going to protect my colleagues from AI-induced intensity. It’s too easy to send someone an AI-generated artefact, without considering its downstream impact.
• To be clear, sometimes speed is necessary to achieve an outcome.
• Speed of outputs, however, isn’t always the bottleneck for a meaningful outcome, or a reliable proxy for productivity. 

21 more annotations...

• In February, Jack Dorsey laid off more than 4,000 people—nearly half of Block’s workforce. “Intelligence tools have changed what it means to build and run a company,”
• Micha Kaufman of Fiverr also made headlines for his stark predictions about AI. In a blunt letter to employees, he wrote, “[H]ere is the unpleasant truth: AI is coming for your jobs
• There’s evidence that companies may follow some version of Dorsey’s path, adopting AI tools with automation and cost-cutting in mind. Goldman Sachs bankers expect the total headcount across their investment banking clients, which span industries, to drop by around 11% on average over the next three years because of AI. And the 2025 Indeed Workforce Insights Report, which surveyed some 80,000 workers across eight countries, found that the time saved with AI was mostly redirected into doing “more of same tasks” or was absorbed by other projects.
• e believe that automation and augmentation will generate different yet predictable behavioral dynamics and associated performance outcomes. Specifically, automation strategies will show early gains relative to the deeper investment required for augmentation, but that augmentation will perform better in the long run.
• As leaders roll out AI strategies, they aren’t simply introducing new tools; they are signaling whether employees have a future in that organization.
• Perceptions of strategic intent are signals that travel fast. When people sense that AI is being deployed to ultimately replace them rather than empower them, their behavior shifts.
• To understand how employees perceive their companies’ intentions around AI, automation, and augmentation, we surveyed 1,294 full-time desk workers
• Notably, only 44% of respondents said their organization had formally announced any AI plans at all. Overall, approximately 62% reported they believe that their organization incorporates AI to augment their capabilities. Thirty-four percent of employees believe their organization is using AI to automate work; 4% remain unsure.
• n some industries (such as retail and professional services) about 40-50% of employees suspect that adopting AI is ultimately undercutting their own job security.
• Other research shows that leaders are vastly overestimating employee enthusiasm for AI. A recent survey found that while 76% of executives believed employees were enthusiastic about AI adoption, just 31% of individual contributors agreed
• 81% of senior leaders think their organization is all-in on augmentation, while those who are closest to the frontlines are more skeptical. At the individual contributor level, 53% perceive augmentation and 40% suspect automation.
• Based on our prior collective research, we believe that employee perceptions should be a central factor in developing a company’s AI strategy,

• 1. The well-being lever.

   

  A leading concern for workers is that AI will lead to layoffs. In our survey, approximately 60% expressed concern about job displacement, with 32% expressing moderate to high concern.

• 2. Workflow integration lever.

   

  Employees are told to “use AI” without clear guidance on why or how it will improve their work. When workers aren’t empowered to take control of this process—as we call being “pilots” of AI—they default to following instructions with limited conviction, like passengers along for the ride.

• 3. The talent pipeline lever.

   

  Long-term, cost-cutting AI strategies can hollow out the junior talent pool. Entry-level white-collar roles are where future leaders build judgment, networks, and expertise.
• Employee perception is also a useful lens for leaders because of how the adoption of new general purpose technologies play out. Integrating a new technology into how we work takes time, as organizations have to invest in new processes, workforce training, data infrastructures, and management practices before productivity gains materialize.
• h “the Productivity J-curve” because of the initial dip in productivity prior to the sharp rise.
• As such, it’s a useful lens for thinking about how AI investments—and the automation vs. augmentation choice—might play out and show up in the company results.
• At its most basic level, automation substitutes human labor with AI in relatively well-specified tasks
• Augmentation, on the other hand, requires deep organizational transformation and learning-by-doing, as well as job redesign and the development of effective human-AI coordination routines—a longer and deeper initial dip.
• The J-curve logic, however, also suggests that augmentation holds greater long-run potential: once complementary investments are absorbed and new socio-technical routines stabilize, performance rises to reflect not just efficiency gains, but a shift in the organization’s productive frontier
• Augmentation is about inventing the future rather than automating the past
• Whether AI is deployed for automation or augmentation, it sends powerful signals that shape employee expectations, feelings, and actions.

• The Automation Path: A Six-Phase Decline

   

  The automation path generates a negative organizational trajectory that undermines long-term performance. When organizations signal that AI is primarily a tool for cost-cutting and workforce reduction, employees rationally disengage, trust erodes, and effort shifts toward short-term self-preservation rather than value creation.

• Phase 1: Early resistance to AI adoption.

   

  When employees sense that AI is being introduced without a genuine commitment to their future—and may be a precursor to layoffs—they naturally begin to resist its broader integration into the workplace.

• Phase 2: Layoffs undermine well-being and productivity.

   

  As redundancies begin, workplace well-being drops sharply. Research one of us contributed to (Jan), found that happy workers are roughly 13% more productive.

• Phase 3: Leaner teams become overburdened—and “workslop” rises.

   

  Smaller teams paired with lower morale result in overextended employees. Many turn to AI to fill gaps in their workflows, often without training or context. This accelerates “workslop,” which undermines efficiency rather than enhancing it

• Phase 4: Rising attrition as employees seek stability.

   

  Declining well-being and rising overload lead more employees to seek opportunities elsewhere.

• Phase 5: Declining employer brand and talent attraction.

   

  As layoffs, disengagement, and churn accumulate, the organization’s reputation as an employer suffers.

• Phase 6: Erosion of leadership pipelines and culture.

   

  In the space of a few years, a deeper consequence emerges: the loss of internal leadership pipelines

29 more annotations...

• a major discontent growing in large companies across the globe. According to the findings, 54 percent of workers reported avoiding their company’s in-house AI tools in order to complete tasks themselves. A full third of workers reported never using AI at all.
• While 61 percent of executives surveyed said they trust the tech for complex, “business-critical” decisions, only nine percent of workers said the same.
• A further 88 percent of corporate bigwigs expressed confidence that the AI tools they forced on their workers were adequate — while only 21 percent of workers agreed.
• 81 percent of executives think their AI deployments have “significantly improved productivity,” their workers are actually wasting eight hours per week cleaning up after AI’s messes, which is the equivalent of 51 work days a year.
• “AI didn’t deliver,” Johns Hopkins economist Steve Hanke told Fortune about the findings. “Welcome to the real world. Forget the AI bubble.
• “Productivity, by the way, it was weak,” Hanke continued. “If AI delivered, productivity would be way up. You listen to these Silicon Valley guys and they say we’re gonna have GDP going to 5 percent [or] 6 percent. Productivity is gonna go up to six. It’s just not happening.”
• MIT study found that 95 percent of AI deployments in the workplace had failed to generate the expected return on investment.

5 more annotations...

• Ce sont donc les agents IA, et plus particulièrement la reconfiguration des processus, entre agents et humains, qui est porteur de cet objectif de ROI.
• Un seul projet agentique sur dix irait actuellement en production. Certains disent même qu'une véritable IA agentique n'est pas encore pour demain.
• Un LLM qui rédige un email, c'est un outil. Un agent qui résilie un contrat, qui refuse un crédit, qui écarte un candidat, qui fixe un dosage médical, c'est un décideur. Et les décideurs, dans nos sociétés, ont des comptes à rendre et engagent la responsabilité de l'entreprise, via une chaîne de hiérarchie.

• Amazon d'abord. L'affaire est devenue la référence mondiale sur les biais algorithmiques. Elle s'est produite en interne, sans jamais aller jusqu'au procès de candidats rejetés.

  L'outil de tri de CV entraîné sur dix ans de recrutements tech, majoritairement masculins, avait appris à pénaliser les profils féminins avec une efficacité remarquable
• La responsabilité du choix des données d'entraînement appartient à l'entreprise, pas au fournisseur du modèle.
• Le véhicule avait correctement identifié la piétonne, mais l'avait classée comme objet non prioritaire, en raison d'un paramètre de sensibilité mal calibré pour réduire les faux positifs
• Surveiller les écrans de monitoring qui affichent en temps réel ce que le système "voit" et notamment la classification des objets, les trajectoires prévues, les décisions imminentes du logiciel.
• Elle avait confié à un algorithme la décision de rupture de contrat avec des livreurs, sur la base de données de géolocalisation, sans intervention humaine réelle. L'article 22 du RGPD interdit les décisions entièrement automatisées ayant un impact significatif sur les personnes
• La compagnie avait tenté de plaider que son chatbot était une « entité distincte », responsable de ses propres hallucinations. En l'occurrence, le-dit Chatbot, qui était en production et accessible aux clients, avait inventé de toutes pièces une politique de remboursement et l'avait  communiquée à un passager en deuil.
• Le tribunal a été lapidaire : une entreprise est responsable de toutes les informations diffusées sur ses canaux, qu'elles proviennent d'un humain ou d'un agent automatisé.
• Et avec l'entrée en vigueur progressive de la directive européenne sur la responsabilité du fait des produits défectueux (PLD), le régime va encore se durcir. Les logiciels d'IA sont désormais classifiés comme des produits. Une victime n'aura plus à décortiquer l'algorithme pour obtenir réparation. Il lui suffira de démontrer le dommage et le défaut du système.
• D'un côté, les pitchs commerciaux : « notre agent gère vos processus de bout en bout, prend les décisions en temps réel, optimise sans intervention humaine ».
  De l'autre, les conditions générales : « le fournisseur décline toute responsabilité quant aux décisions prises par le système ».
• Gartner estimait début 2026 que les entreprises devraient provisionner collectivement plus de 10 milliards de dollars d'ici mi-2026 pour couvrir amendes, remédiation et contentieux liés à des décisions automatisées illégales.
• Aux Etats-Unis, on commence à voir émerger des fonctions qui n'existaient pas il y a trois ans : AI Auditor, Decision Reviewer, Algorithm Compliance Officer. Dans les secteurs régulés, santé, finance, services publics, ces rôles vont devenir incontournables
• A la DSI on est aujourd'hui engagé dans l'explicabilité de fonctionnement des plateformes Cloud et surtout de l'UX. C'est une vision uniquement technique. Ce que voit GreenSI apparaître c'est l'explicabilité IA, qui est une nouvelle traçabilité des décisions

• En synthèse de ces 4 cas réels, on retiendra que la responsabilité de l'IA est une responsabilité d'entreprise, sans détour ni sous-traitance possible.

• On peut quand même structurer l'approche, sans se faire peur, avec trois réflexes à intégrer dans chaque déploiement d'agent autonome :

  • Auditer avant de déployer : les données d'entraînement, les seuils de sensibilité, les cas limites. Ce travail n'est pas optionnel, il constitue la première ligne de défense juridique.
  • Tracer toutes les décisions à impact : ruptures de contrat, refus, alertes, recommandations médicales. La reconstruction a posteriori d'une décision IA sans logs est illusoire.. C'est la boîte noire.
  • Nommer un humain dans la boucle : un réel point de contrôle. La supervision symbolique ne trompe plus personne, ni les juges ni les régulateurs.

15 more annotations...

• espite the transformative effect of AI, customers are asking Deloitte: Where is the ROI from our AI investments? When will we see the efficiency gains?  
• hey are stuck in a “process trap” whereby they are inserting AI into existing processes developed for deterministic technology, whereas AI is probabilistic.  
• “I won’t suggest that human-centered design is anything new” – it’s been around since the industrial revolution.  

• employers have lots of choices and decisions to make when it comes to redesigning work.  

   
• how do we think about a business use for humans?” asked Szpaichler. Just because you can automate something does not mean you should.  
• Deloitte advises that organizations set a clear boundary for where you need humans not just in the loop, but in full control
• there’s nothing that’s getting asked of a HR function that hasn’t been asked of them before. The difference is the speed that they’re expected to respond to it, and the spotlight they’re under.” 

5 more annotations...

• It’s not just Meta, though. Nvidia CEO Jensen Huang said in the All-In Podcast that if a $500k engineer spent less than $250k a year in tokens, he’d be “deeply alarmed.”
• Token usage is, by all accounts, a new status symbol. Engineers at the top labs boast about token numbers like a bad coder would about lines of code: look how excessive I am. They are “tokenmaxxing,”
• The more you spend, the more they think, and, presumably, the better the output you get, and thus the more benefit you create for your company.
• Some ask, dumbfounded: What do you think your employees will do if you take this simple, hackable metric as proof of work? Well, hack it. Duh. And so you have guys instructing their bots to run loops with the sole goal to burn tokens or to “be as token inefficient as possible.”
• The idea that “the more tokens you use, the more value you provide” gets backward not only the economic incentives, but also the relationship between thinking and language
• And I won’t ever bother explaining to myself what that thing is with all the words I’d need if I were to explain it to someone else, which might amount to one or two hundred, or maybe five thousand, depending on the topic.
• I’d say that thinking with words is like moving from point A to point B, whereas thinking in sensations is akin to teleporting from point A to point B. You should try it. (That is, by the way, my best take on what I believe “thinking in latent space” must feel like for AI models; sensations rather than words. What is it like to be a pre-trained model, right?)
• The whole thing happens pre-linguistically, much faster than it would take me to put it into words. Only later, I decide whether I want to polish it with the structural superiority of language.
• . But words are a lossy compression of something richer. More concrete but less precise; more transferable but less detailed.
• “language is primarily a tool for communication rather than thought.” It “only reflects, rather than gives rise to, the signature sophistication of human cognition.”
• AI models have the infrastructure to process information at a pre-linguistic level as well, in the “latent space.” It’s the AI labs who force them, through the process of post-training, to move through a taxing mire of sequential symbol generation when they apply inference-time techniques,
• The real solution would be to build an AI system that had the intelligence and “reflection” skills of post-trained models,
• You may be wondering why this “inference-time thinking” matters at all. Well, it matters because AI models that generate vast quantities of tokens at inference time with the sole goal of “thinking”—that is, improving benchmark scores—are the sole reason why companies like Meta are reaching two-digit trillion token spend on Claude. Thinking in tokens is, thus, not the celebrated breakthrough they made it to be last year.
• It is, actually, a prosthesis, a scaffolding apparatus that the AI labs used to erect a taller building and then, when they realized the building would immediately collapse if they removed it—going back to pre-training is not an option—they started to convince us—and themselves—that the scaffolding was actually load-bearing architectural elements:

• It’s just not economically feasible to overhaul an entire industry just because it’s resting on top of a fallacy. Won’t somebody please think of the shareholders?

  So, no, I won’t get between you guys and your Token Legend status.
• While the labs kept amassing funding and building datacenters, these outcasts were trying to undo the absurdity of forcing AI models to think in tokens. “We will never achieve true machine intelligence if we keep using tokens to encode thought,”
• They don’t ask: how do I efficiently scale all these grains of sand to build my cathedral? Or: How do I pretend that this scaffolding belongs to the cathedral? No, they make sure to invent the brick and the column and the arch and the vault.
• Indeed, Meta’s FAIR lab—Yann LeCun’s group—published “Coconut” (Chain of Continuous Thought) in late 2024, a model that reasons through continuous space instead of decoding into tokens

• Naturally, having used JEPA to predict the caricature that Meta would become with the precision that only a sensation-level thinker can achieve, LeCun left in late 2025.
• LeCun has been calling LLMs “a dead end” for years, but Meta went all in in a sort of FOMO-fueled fickleness that’s idiosyncratic of Mark Zuckerberg.
• To me, Meta has been out of the race for a while, but this delay, together with the constant restructuring and LeCun’s departure, was the definitive sign.
• Did Meta just spend 60 trillion tokens last month on Anthropic models to distill from Claude’s reasoning traces the data they needed to bootstrap this new release?
• Here’s how they could´ve achieved it: Meta spins up a bunch of accounts (not fake, as were those from the Chinese labs) and then uses those accounts to bombard a stronger model (Claude) with carefully crafted prompts designed to extract its reasoning patterns.
• But even if illegality is off the table, if this is true, it’s a stain on Meta’s efforts to catch up with the top labs. Whereas Anthropic and OpenAI have been working hard to achieve a much better pre-trained model onto which to build the next generation of agents, Meta is out there, with more compute than it can use, struggling to milk Claude, only to fail to beat it on the most important benchmarks.
• The thesis I had in mind—that it’s time for AI models to stop thinking with words—needs to be updated. If my suspicions are correct, the true thesis is this: It might be a waste of compute that current AI models generate so many tokens when they think, but if they didn’t—if they could think with sensations rather than words—then Meta couldn’t have trained Muse Spark.

23 more annotations...