• ​The text outlines a new landscape where HR’s value shifts from “owning processes” to becoming architects of human capability systems.
• Most of the HR models in use today were designed for industrial-era organisations built around human labour, hierarchical authority and slow information flows.
• We are entering the Fusion Era, where the workforce is no longer exclusively human, and intelligence is embedded directly into systems
• AI-centric organisations will not need HR departments as process managers but architects of human capability systems instead.
• In traditional organisations, information systems record activity after decisions are made, whereas in AI‑centric organisations, intelligence is embedded inside the workflow itself and systems do not simply store information, they interpret it, analyse it and when required, act on it.
• • Decision support becomes continuous rather than occasional.
      
  • Workflows become algorithmically orchestrated.
      
  • Employees interact directly with systems rather than through administrative intermediaries.
• In an AI‑centric organisation, the system executes the process, and the resulting implication is clear: if HR continues to define its value through owning processes, it is building its future on disappearing work.

• These capabilities are not there to support HR; they bypass it.

      
• • Human Judgement
   
      

  Strategic direction, ethical choices and organisational identity remain human responsibilities requiring context, accountability and moral reasoning. AI may well inform these choices, but it does not replace them.
• • AI‑Assisted Decisions
   
      

  In some operations, systems generate insights while humans validate the outcome. In this category we can include recruitment screening, workforce risk alerts, compensation benchmarks and learning recommendations. The system performs the analysis, the human makes the judgement.

      
   
• • Autonomous System Decisions
   
      

  Some decisions can become fully automated once reliability of the process has been established. Scheduling, workflow routing and resource allocation will increasingly operate without human intervention. The system does not advise, it acts.
• without governance they may hallucinate, misinterpret data or trigger unintended actions across enterprise workflows.

• Trust requires governance, and that governance can only come from:

      

  • clear access and control over who builds and deploys agents 
  • guardrails that restrict system behaviour 
  • decision logic audit mechanisms 
  • strong data governance 
  • the lifecycle management of AI systems 
  • usage cost monitoring  
  • continuous system oversight

• The principle is simple: Trust but verify.

      
• . The design question, therefore, instead of ‘How many employees do we need?’ becomes ‘What combination of human and machine capability produces the best outcome?’ In reality, hiring a human becomes a design decision.
• Developing and rolling out new technology in the middle of a product trajectory is a very expensive process, not only in execution but in marketing as well. It’s not difficult to envision some of the long-established ‘legacy’ players struggling to change course in a market that has become suddenly dynamic due to growth of single-module applications and the greater ease in linking those modules together when needed.

• future lies in having a role in the architecture of human capability systems viz:

      
   
  • human‑AI work design 
      
  • workforce capability architecture 
      
  • decision governance 
      
  • workforce ethics and trust
      
  • legislative compliance

15 more annotations...

• In Singapore, AI engineer hiring has grown by 89% over the past three years, making it the fastest-growing job in the country.
• LinkedIn data show an 82% year-on-year increase in members reporting feeling “overwhelmed” by the rapid pace of change.
• New research further underscores a growing “hiring disconnect” across the region. While professional activity is high—with 58% of employees in Singapore and 72% of employees in India actively seeking new roles in 2026—competition is fierce.
• traditional hiring methods can no longer keep pace with the rapid shift in skill requirements
• introducing the “Superworker Organization” framework. Bersin argued that we are currently witnessing a fundamental shift from AI as a personal productivity tool to AI as an enterprise transformation technology.
• According to Bersin’s latest research, roughly 40% of current HR activities can now be automated. This does not imply mass unemployment; rather, it signals the birth of the “super job.”
• the time saved on administrative tasks is reinvented into higher-value activities such as consulting, advising and coaching.
• He noted that the most profitable organizations in 2026 are those that have mastered the art of internal reskilling and dynamic redeployment, treating talent as a fluid resource.
• he “new world of work” demands a shift toward a “super agent” architecture, where various AI tools—from recruiting assistants to onboarding bots—are integrated into seamless workflows.
• is helping early adopters like Wipro and Siemens save over four hours per role by reviewing 62% fewer profiles while simultaneously seeing a 69% improvement in InMail acceptance rates.
• Bersin introduced the concept of the “super manager”—leaders who move away from top-down edicts and instead facilitate a culture of creativity and pioneering.
• The gap between exponential technological change and slow organizational adaptation can only be closed by managers who empower their teams to experiment and iterate without waiting for centralized IT approval.
• The transition to an AI-driven organization is not merely a technical upgrade; it is a leadership challenge.

11 more annotations...

• summoned a large group of engineers to a meeting addressing recent outages plaguing its online retail business, some of them related to AI coding tools.
• described the “trend of incidents” as characterized by a “high blast radius” and “Gen-AI assisted changes.”
• the availability of the site and related infrastructure has not been good recently,
• leaving customers unable to make orders. In the aftermath, the company blamed a botched “software code deployment.”
• two separate outages were caused after engineers allowed the company’s in-house AI coding tool to make disastrous changes
• ramed these blunders as an issue related to its protocols around AI usage and “user access control,” rather than an AI autonomy issue
• Junior and mid-level engineers will now require more senior engineers to sign off any AI-assisted changes
• frequently allow errors through and sometimes struggle to follow instructions, meaning they can take actions that a user never intended. 
• But Amazon’s renewed focus on implementing more human oversight comes as it’s fired hundreds of workers from its cloud computing division and as it targets laying off 30,000 employees across its corporate workforce overall

• In sum: more coding with more AI with more human oversight, but fewer humans. We’ll see how that works out.

8 more annotations...

• Most people assume that admitting you could be wrong costs you credibility.
• No. You trust it more. You trust it more than if it said nothing, and far more than if it claimed to always be correct
• This runs against a deeply held assumption in most workplaces — that admitting you might be wrong undermines your credibility.
• People around them learn, quickly, what kind of information is welcome and what isn't. They filter. They hedge. They wait and see.
• . It's an operational problem with direct costs. Decisions get made on incomplete information.
• At Toyota, this insight is baked into the production system itself. Any team member can pull an andon cord to flag a problem — and the expectation is that they will.
• The organizations that surface problems fastest are the ones that fix them fastest.
• when we share something vulnerable, we tend to assume others will think less of us.
• ncludes “acknowledge your own fallibility” in her recommendations for leaders who want teams that speak up and improve
• When a leader says, “I may be missing something — what are you seeing?” it signals that being wrong is a normal part of doing work
• The leaders who project certainty at all times aren't building confidence. They're building distance
• In a world where almost every leader defaults to projecting certainty, the one who signals fallibility is doing something genuinely unusual. And unusual signals get noticed. They're memorable precisely because they're rare.
• A disclaimer isn't a weakness. It's a trust signal. It says: a human is involved here
• The upside is harder to picture because it's never a single dramatic moment.
• That's the hard part. The payoff is real, but it's distributed over time and across decisions. The risk feels concentrated in one visible moment.

13 more annotations...

• Anthropic has introduced a new metric for enterprises to determine where AI is making an impact.
• In the report, the vendor used an "Observed Exposure" metric to look at three data sources.
• They are whether a large language model (LLM) such as the vendor's Claude model can perform a task twice as fast as a human; how Claude is actually being used in work-related settings, and how job descriptions show AI’s use in specific roles. Using the specific role metric, Anthropic found that actual AI use in the workplace is lower than thought.
• there has been no systematic increase in unemployment among workers, such as older professionals, women, more educated individuals, and high-paid employees, since 2022
• , hiring of younger workers has decreased, and the job roles most exposed to AI, such as computer programmers, customer service representatives, and those for which writing, reading and research can be automated

• humans are still not using AI technology to its full potential, according to the report.
• These events have affected a job market in which many feel dread that an LLM will snatch their jobs.
• the Anthropic report echoes what some experts have concluded: it's too early to blame AI for layoffs, since it's still unclear what effect the technology is having on jobs.
• it's difficult to trust any quantitative measure of impacts on labor,"
• "Some employers may be using the notion of AI displacement of human workers to help justify cuts they would've made anyway,"
• "We need more and more nuanced metrics," he said. "This one will likely encourage the development of new approaches."

9 more annotations...

• But while many of these pilots have succeeded individually—they’ve saved time and money, made processes more efficient—those gains haven’t scaled across the organization.
• to study how firms move beyond simple automation toward agentic workflows that change how companies create value.
• These participants were enthusiastic adopters of AI tools, not laggards.

• 7 Frictions Stalling the Last Mile

   

  Seven specific frictions tend to stall this transition.

• The Proliferation of Pilots

   

  Many organizations find themselves “pilot-rich but transformation-poor.” For example, a global investment bank recently documented more than 250 applications that connected LLMs to its enterprise systems.
• Despite these impressive figures, leaders at the summit reported that what works in a localized demonstration rarely becomes the standard

• The Productivity Gap

   

  While productivity gains are often visible at the individual level, they frequently fail to manifest on the balance sheet. A global payments network reported that more than 99% of its employees actively use copilot
• These gains remain trapped inside individual workflows unless leadership intentionally redesigns roles and budgets to capture the reclaimed time. This occurs because saved time is often re-absorbed into low-value activities—like more internal meetings or unnecessary emails—rather than being structurally harvested through role reclassification or a mandate and clearance to shift time toward higher-value strategic work.

• The Burden of Process Debt

   

  AI often serves as a diagnostic tool that exposes brittle, exception-ridden processes. At many companies, these are accumulated over decades of acquisitions, but most organizations suffer from degrees of this problem
• In these environments, the challenge is the need to re-architect the workflow itself so that AI can operate reliably.

• The Identity Problem of Tribal Knowledge

   

  A significant barrier to the last mile is the reliance on tacit knowledge that lives in the heads of long-tenured employees.
• An engineering consultancy framed this as an identity problem rather than a reskilling issue

• Governance in an Agentic World

   

  As organizations move toward agents that take action—such as updating systems or coordinating across applications—traditional governance becomes a bottleneck.
• A global bank noted an emerging accountability gap where human-in-the-loop controls, which work for isolated cases, collapse under multi-agent architectures.

• Architectural Complexity

   

  Most enterprises operate a patchwork of AI capabilities across multiple hyperscalers (cloud computing service providers at industrial scale) and application stacks.
• A global apparel company described the months required to make agents across SAP, Microsoft, and Google environments communicate reliably.

• The Efficiency Trap

   

  The initial framing of AI as a cost-reduction tool has become a constraint for many firms. Several organizations compared the early positioning of AI to a new form of offshoring, which triggered defensive behavior from middle management and narrowed the scope of the C-suite’s transformation ambition.
• Despite these frictions, the organizations making the most significant progress are converging on a shared operating-model blueprint characterized by several strategic shift
• Rather than a menu of optional tactics, these seven strategic pillars constitute a unified operating model.

• Clean-Sheet Process Redesign

   

  Leading firms have moved beyond bolting AI onto legacy workflows. Instead, they treat the technology as a trigger for clean-sheet thinking. Clean-sheet redesign involves discarding legacy workflows entirely and asking: “If we were building this process today with modern AI agents, how would we do it?”

• Strategic Knowledge Capture

   

  Organizations that move quickly treat tribal knowledge as a strategic asset. In an AI context, this means capturing the “pre-harvest” context (rules and logic required before a task begins) and “execution” context (the nuanced data passed between agents during a task).

   

  Several firms are systematically identifying expertise-dependent processes and pairing senior experts with designers to externalize their judgment.

• Managing the Digital Workforce

   

  Model governance is no longer sufficient; the new frontier is agent governance. Forward-looking organizations are designing control planes that define who can create agents, what actions those agents can take, and where accountability sits when things go wrong.

• Role Redesign and Career Pathing

   

  As AI absorbs more execution-oriented tasks, human roles must shift toward design, orchestration, and interpretation. Without this explicit redesign, AI adoption can feel like the erosion of a career rather than its evolution. This shift increases the demand for leaders who will explicitly guide and verify AI output rather than doing the manual work themselves. A new job description looks for “learning agility” and “domain depth” while ensuring clear ownership of “agent orchestration.”

21 more annotations...

• Très honnêtement, je pense que nous sommes largement plus proches du verre à moitié vide, voire d’un verre vide avec 2 ou 3 gouttes dedans que l’on nous exhibe comme une preuve formelle d’un raz-de-marée à venir.
• nombreux sont les techno-prophètes sur YouTube ou LinkedIn qui vous promettent de gagner 5 h par jour grâce à l’IA,
• les éditeurs et promoteurs de l’IA qui, dans leur empressement à monétiser une technologie très coûteuse, font fausse route.

• Ces six boucles partagent un point commun : elles évitent toutes la question centrale. Sortir de l’impasse exige d’entrer dans ce que Scaman appelle la Terra Nova : un territoire où l’on cesse de débattre du remplacement ou du refus pour expérimenter ce que l’IA permet de nouveau.

   

      
• Ce n’est pas « L’IA est-elle bonne ou mauvaise ? » qui devrait nous occuper, mais « Comment l’IA peut-elle s’intégrer utilement dans le quotidien professionnel sans générer de nuisances ? »
• Nous sommes encore dans l’équivalent de 1995 pour l’internet, quand la grande majorité des décideurs ne voyaient pas en quoi le « réseau des réseaux » allait transformer leur activité
• Dans l’ensemble, en termes de création et de destruction d’emplois, nous ne trouvons aucune différence significative entre les entreprises qui déclarent utiliser l’IA et celles qui ne l’utilisent pas
• cette autre étude de l’éditeur de Claude qui nous démontre que, si l’impact de l’IA est très faible sur les emplois, c’est parce qu’elle est sous-exploitée :
• une étude de plus ou de moins ne fera pas la différence, car les raisons de la défiance des salariés envers l’IA sont irrationnelles
• Ce qui est certain, c’est que la surenchère technologique des éditeurs ne répond pas à la question principale : comment faire monter en compétences des utilisateurs qui ne se sentent pas concernés, voire qui résistent activement ?
• les éditeurs ont décidé d’imposer l’IA de force, avec des discours anxiogènes du type « Vous allez être remplacé par une IA ou par quelqu’un qui utilise l’IA« , il y a eu un phénomène naturel de rejet de la part des utilisateurs potentiels.
• l’IA générative n’est utilisée à son plein potentiel que par une infime minorité de la population (autour de 0,1%)
• l’utilisation de l’IA générative nuit à la qualité de l’apprentissage : « Successfully performing a task with GenAI does not automatically lead to learning »
• . L’urgence n’est pas d’améliorer les modèles, mais de mieux marketer l’IA, de la repackager pour la « vendre » de façon plus efficace aux utilisateurs potentiels.
• La première ère est celle des grands systèmes dans les années 1960-80 avec leurs scripts et batchs. Ces systèmes conçus par IBM exploitaient des logiciels avec une approche procédurale : des instructions, développées par des informaticiens, qui sont exécutées les unes à la suite des autres.
• La deuxième ère est celle des PC dans les années 1980-2000 avec tous leurs logiciels installés sur des ordinateurs individuels.
• La troisième ère est celle du cloud computing dans les années 2000-2020 et des applications en ligne. Une véritable révolution pour les DSI, qui ont pu s’affranchir des contraintes de déploiement et de mise à jour des logiciels, car tous les traitements sont centralisés dans des data centers
• La quatrième ère est celle de l’IA, relancée par les modèles génératifs et les agents intelligents. Là encore, une nouvelle révolution avec un nouveau paradigme : celui de la délégation et de l’orchestration.
• Il n’est plus ici question pour les utilisateurs de chercher une fonctionnalité dans un menu ou d’assimiler le fonctionnement d’une interface, mais d’exprimer un besoin,
• Nous parlons bien ici d’une approche agentique de l’IA, pas seulement de poser des questions à un chatbot et d’attendre une réponse.
• Dans la mesure où l’IA générative et les agents intelligents représentent un nouveau paradigme des services numériques, je milite pour un nouveau discours reposant sur des appellations commerciales incrémentales.
• « IA », « modèles génératifs », « agents intelligents » sont des termes qui désignent des briques technologiques, pas des usages.
• Officiellement, les statistiques d’adoption sont très encourageantes
• Mais dans les faits, les usages réels sont plutôt basiques, comme en témoignent les statistiques fournies directement par OpenAI : les usages de ChatGPT concernent majoritairement des conseils pratiques ou de la recherche d’informations
• es chatbots proposent une interface trop déroutante pour des utilisateurs qui ne savent pas trop quoi faire avec et se contentent de reproduire ce qu’ils connaissent déjà
• prompter efficacement est une compétence non triviale. Savoir formuler une instruction précise, contextualiser une demande, décomposer un problème complexe en étapes séquentielles… nécessite un apprentissage, et cela prend du temps.
• les choses sont en train de changer avec des agents intelligents préconfigurés, qui combinent la puissance et la souplesse des modèles génératifs avec la simplicité de prise en main des interfaces traditionnelles.
• Nous sommes maintenant au tout début d’une nouvelle ère, le web en mode « Action » (« Act« ) : avec l’avènement des IA génératives, les utilisateurs peuvent maintenant faire (quasiment) tout ce qu’ils souhaitent, grâce aux agents intelligents qui vont agir à leur place.

• Les agents intelligents étant potentiellement plus faciles à comprendre, la notion de web agentique peut être simplifiée en « Web4 ».

      
• Puis, avec l’avènement du Web 2.0 dans le milieu des années 2000, sont apparus les premiers blogs d’entreprise, wikis et réseaux sociaux internes, ainsi que les nombreuses applications en ligne
• Idem pour le Web3, avec les nombreuses applications potentielles de la blockchain dans le monde professionnel (décentralisation, traçabilité, smart contracts…), mais qui restait un concept technologique complexe à expliquer et encore plus à comprendre pour le salarié lambda
• Nous nous retrouvons donc aujourd’hui face à un nouveau risque de dispersion avec de nombreux chatbots et services dopés à l’IA générative qui s’ajoutent à un environnement de travail déjà surchargé.
• repenser les environnements de travail, à nouveau, en y intégrant l’IA, non pas comme une couche supplémentaire, mais comme une composante fondamentale de l’architecture des outils
• Le pré-requis est de traduire les sacrosaints cas d’usage (qui n’existent souvent que sur le papier) en outils prêts à l’emploi, donc pas sous la forme de chatbots avec une interface conversationnelle qui rebute plus qu’elle ne facilite.
• ’il existe un standard agentique simple à implémenter et relativement facile à comprendre pour les utilisateurs : les skills (pour les tâches horizontales, transversales aux différentes fonctions de l’entreprise) et les plugins (des packages de skills adaptés aux tâches verticales, donc orientés métier).
• il existe même des outils complémentaires pour aider les modèles dans leur phase d’appropriation du contexte comme Vercept, récemment racheté par Anthropic, qui permettent d’enrichir les agents d’une connaissance fine des pratiques et interfaces de l’organisation.
• Oui je sais, le concept de jumeaux numériques est encore trop avant-gardiste pour les salariés lambda des entreprises lambda. Mais nous y allons petit à petit : les chatbots sont une première étape,

35 more annotations...

• Worse still, NVIDIA’s earnings also mentioned $27bn in cloud commitments — literally paying its customers to rent the chips it sells, heavily suggesting that there isn’t the underlying revenue.
• A day later, CoreWeave posted its Q4 FY2025 earnings, where it posted a loss of 89 cents per share, with $1.57bn in revenue and an operating margin of negative 6% for the quarter.
• it shows that CoreWeave’s underlying business appears to be weakening as it adds capacity, which is the opposite of how a business should run
• CoreWeave, as I’ve been warning about since March 2025, is a time bomb. Its operations are deeply-unprofitable and require massive amounts of capital expenditures
• This suggests a fundamental weakness in the business model of renting out GPUs,
• I think it’s reasonable to ask whether NVIDIA might have sold hundreds of billions of dollars of GPUs that only ever lose money.
• Lots of people have come up with very complex ways of arguing we’re in a “supercycle” or “AI boom” or some such bullshit, so I’m condensing some of these talking points and the ways to counteract them:
• OpenAI had $13.1bn in revenue in 2025! They only lost $8bn
• OpenAI has 900m weekly active users!
• Google Gemini Has 750m-
• Claude Code is changing the world! It’s writing SaaS now! It’s replacing all coders!
• People Doubted Amazon But It Made Lots Of Money In The End!
• AWS Cost A Lot Of Money So We Should Spend So Much Money On AI!
• • OpenAI raised $15bn from Amazon, with $35bn contingent on AGI or an IPO.
  • OpenAI got commitments from SoftBank and NVIDIA, who may or may not have committed to $30bn each, and will be paying in three installments.
• OpenAI’s nonprofit entity, which has a stake in the for-profit OpenAI that’s now worth $180bn, may sell several billions of dollars of its shares to the financial investors, depending on the level of investment demand the for-profit receives in its fundraise, the person said. That would help other OpenAI shareholders avoid additional dilution of the value of their shares following the large equity fundraise.
• I get where the concern comes from, but I don’t think it matches my understanding of how this all works. This only makes sense if new revenue flows into the whole AI ecosystem. If people are not willing to pay for the services that we and others offer, if there’s not new economic value being committed, then the whole thing doesn’t work
• Now, jokes aside, ponzi schemes work entirely through feeding investor money to other investors. OpenAI and AI companies are not a ponzi scheme. There’s real revenues, people are paying it money. Much like NVIDIA isn’t Enron, OpenAI isn’t a ponzi scheme.
• It’s very obvious that neither OpenAI nor its peers have any plan to make any of this work beyond saying “well we’ll just keep making more money,”
• That’s right, by the end of 2026 OpenAI will make as much money as Paypal, by the end of 2027 it’ll make $20bn more than SAP, Visa, and Salesforce, and by the end of 2028 it’ll make more than TSMC, the company that builds all the crap that runs OpenAI’s services.

• It’s just that easy. And all it’ll take is for OpenAI to burn another $230 billion…though I think it’ll need far more than that.
• I doubt these numbers! I think they are very shifty! My own numbers say that OpenAI only made $4.3bn through the end of September, and it spent $8.67bn on inference! Nevertheless, I can still make my point.
• • Per The Information, OpenAI was at $17.5bn in cash and cash equivalents at the end of June 2025. It had just raised $10bn from SoftBank and other investors.
  • OpenAI would raise another $8.3bn on August 1 2025, bringing that cash and equivalents pile to $25.8bn, assuming it remained untouched.
  • OpenAI would raise another $22.5bn from SoftBank on December 31 2025, bringing up the total to $48.3bn.
  • In the second half of the year, OpenAI would (allegedly) make another $8.8bn, which would bring us up to $57.1bn — with a total year loss of either $9bn or $8 bn depending on whether you believe The Information or CNBC. 

• In my mind, it’s far more likely that OpenAI’s losses were in excess of $10bn or even $20bn, especially when you factor in that OpenAI is paying an average of $1.5 million in yearly stock based compensation, per the Wall Street Journal. 

• What I’m trying to get at is that OpenAI (and, for that matter, Anthropic) has spent the last two years increasingly obfuscating the truth through leak after leak to the media. 

  The numbers do not make any sense when you actually put them together, and the reason that these companies continue to do this is that they’re confident that these outlets will never say a thing, or cover for the discrepancies by saying “these are projections!” 

• hief Executive Dario Amodei has spent the last few years massively overstating what LLMs can do in the pursuit of eternal growth. 

  He’s also framed himself as a paragon of wisdom and Anthropic as a bastion of safety and responsibility.
• Anthropic, of course, leaves out one detail: Hegseth said that “...effective immediately, no contractor, supplier, or partner that does business with the United States military may conduct any commercial activity with Anthropic.” If Hegseth follows through, Anthropic’s business will collapse, though Anthropic and its partners are ignoring this statement as a supply chain risk only forbids Anthropic from working with the US government itself.
• The military’s demands were for “all lawful uses,” though I don’t think Anthropic really gives a shit about whether the war in Iran is legal, because if it did it would have shut down the chatbot rather than supported the conflict.
• In reality, Claude is likely being used to go through a bunch of images and to answer questions about particular scenarios. There is very little specialized military training data, and I imagine many of the demands for “full access to powerful AI” have come as a result of Amodei and Altman’s bloviating about the “incredible power of AI.”

• “What’s right,” to be clear, involves allowing Claude to choose who lives or dies and to be used to plan and execute armed conflicts. 
• The real problem will be if Hegseth is serious about the stuff that isn’t legally within his power — namely limiting contractors, suppliers or partners from working with Anthropic entirely.
• If Hegseth and the administration were to truly want to punish Anthropic, they could put pressure on Amazon, Microsoft and Google to cut off Anthropic, which would cut it off from its entire compute operation — and yes, all three of them do business with the US military,

• On Friday afternoon, Anthropic learned that the Pentagon still wanted to use the company’s AI to analyze bulk data collected from Americans. That could include information such as the questions you ask your favorite chatbot, your Google search history, your GPS-tracked movements, and your credit-card transactions, all of which could be cross-referenced with other details about your life. Anthropic’s leadership told Hegseth’s team that was a bridge too far, and the deal fell apart.
• Anthropic had not argued that such weapons should not exist. To the contrary, the company had offered to work directly with the Pentagon to improve their reliability.
• So, let’s be clear: Anthropic wants to help the military make more accurate kill drones, and in fact loves them.
• Anthropic is totally fine with the US military using Claude for anything else, even though hallucinations are an inevitable result of using a Large Language Model. 
• AI safety and wide distribution of benefits are the core of our mission. Two of our most important safety principles are prohibitions on domestic mass surveillance and human responsibility for the use of force, including for autonomous weapon systems.  The DoW agrees with these principles, reflects them in law and policy, and we put them into our agreement.
• One source familiar with the Pentagon’s negotiations with AI companies confirmed that OpenAI’s deal is much softer than the one Anthropic was pushing for, thanks largely to three words: “any lawful use.”
• Altman approving of non-domestic AI surveillance, saying he “didn’t like it” but “accepted it,” echoing The Day Today’s Peter O'Hanraha-hanrahan. 
• A late-evening story from Axios on Monday reported that “OpenAI and the Pentagon have agreed to strengthen their recently agreed contract, following widespread backlash that domestic mass surveillance was still a real risk under the deal — though the language has not been formally signed.”
• This announcement gives OpenAI the air cover to pretend it got exactly the same deal as Anthropic, even though those nasty little words allow the DoD to do just about anything it wants. Oh, it wasn’t deliberate surveillance, we just looked up whether some people had said stuff about the administration.
• The funniest outcome of this chaos is that many people are very, very angry at Sam Altman and OpenAI, assuming that ChatGPT was somehow used in the conflict in Iran, and that Amodei and Anthropic somehow took a stand against a war it used as a means of generating revenue.
• And in both cases, these men are both enthusiastic to work with a part of the government calling itself the Department of War. Both of them are willing and able to provide technology that will surveil or kill people, and while Amodei may have blushed at something to do with autonomous weapons or domestic surveillance, neither appear to have an issue with the actual harms that their models perpetuate.
• ltman vaguely cares about the things that an LLM “isn’t very good at,” Once Claude is used to choose places to bomb and people to kill, suddenly Anthropic cares that “frontier AI systems are simply not reliable enough,”
• Altman and Amodei want it both ways. They want to be pop culture icons that go on Jimmy Fallon and thought leaders who tell ghost stories about indeterminately-powerful software they sell through deceit and embellishment.
• “believe deeply in the existential importance of using AI to defend the United States and other democracies, and to defeat our autocratic adversaries.”
• We’re barely two weeks divorced from the agonizing press around Amanda Askell, Anthropic’s “resident philosopher,” whose job, per the Wall Street Journal, is to “teach Claude how to be good.” There are no mentions in any story I can find about what she might teach Claude about what targets are considered fair game in military combat. 

45 more annotations...

• Organizations measure AI adoption because they can’t measure AI value. As Christopher S. Penn puts it: “If you don’t know the ROI of what you’re doing today, you cannot calculate the ROI of AI’s impact on it.”
• According to Worklytics’ 2025 survey, 74% of organizations say better metrics are critical. Only 17% feel effective at measuring true value.
• Wrong metrics don’t just fail to measure success. They actively incentivize behaviors that destroy value.
• 41% of employees now receive what Harvard Business Review calls “workslop”, AI-generated content that looks professional but contains no substance. Each instance costs 2 hours to decode, verify, redo.
• 32% of workslop recipients are less likely to work with the sender again. Metrics optimized for individual productivity destroy organizational collaboration.

• Most organizations aren’t building motorways. They’re measuring how fast employees pedal.

• AI coding assistants increase developer output, but not company productivity. Faros AI tracks it precisely: “Downstream bottlenecks absorb the value.”
• Developers ship code 35% faster, but QA can’t keep up and features pile up in review queues
• Meanwhile, 44% of marketers now use AI for content summarization. The loop closes: AI generates content humans can’t process, so organizations deploy more AI to summarize what AI created.
• The primary skill becomes knowing what to ignore. That’s not productivity. That’s triage.
• Only 5% of companies achieve what Boston Consulting Group calls “future-built” statu
• Harvard and Mayo Clinic researchers found that combining physician intuition with algorithmic analysis cuts hospital readmissions by 26%;
• Most organizations can’t define what AI success looks like because they’ve never clearly defined what the underlying work was supposed to accomplish
• What the 5% share is less a playbook than a mindset. They ask, seriously, which problems actually warrant deploying AI, and which ones just seem convenient.
• HBR research distinguishes “pilots” from “passengers”: pilots use AI with intent, to extend what they’re capable of; passengers use it to offload work they didn’t want to do. Pilots use AI 75% more, and to better effect.

• The 5% who see transformative returns built the motorway before they bought the supercar.

14 more annotations...