This week in the global AI scene, a debate over 'trust' in business, steady progress toward practical agents, and disputes over research direction all moved at once. Behind the headline-grabbing major releases, voices questioning the industry's structure and ideology stood out.
Palantir's Karp Calls the AI Industry 'Marxist'
Speaking just after the company posted $1 billion in profit for the quarter, Palantir CEO Alex Karp renewed his warnings about AI frontier labs. He repeatedly said the frontier labs are 'not trustworthy for enterprises,' and described the industry's structure as 'Marxist.'
Karp's argument has been consistent: he is skeptical of entrusting corporate secrets and control to the labs building frontier models. For Palantir, whose track record lies in enterprise data analytics, AI trust and control are themes that go to the heart of its sales pitch. While there is an element of self-promotion in his remarks, they can also be read as reflecting a reality: enterprises are rethinking not so much the models themselves as whom to entrust with operations and control.
Microsoft's 'Skill Recorder' — Record Your PC Workflows and Hand Them to Copilot
Microsoft has released a free desktop app called 'Skill Recorder' that lets you record tasks on your PC and helps AI reproduce those same tasks. Skills created with the app can reportedly be used in the agent services 'Microsoft Copilot Cowork,' 'Microsoft Scout,' and 'Copilot Studio.'
The approach — recording on-screen operations and having an agent replay them — points to the possibility of automating routine work without programming. At this point it is aimed mainly at macOS, with Windows support on the way. The trend of agents learning human workflows looks set to lower the barrier to automation, which has so far been code-centric, and open it up to non-engineers.
Is AI Research's Next Hard Problem 'World Models Over Giant LLMs'?
On technical forums, a debate drew attention: does AI research need world models more than ever larger LLMs? The argument is that generating hypotheses is no longer the main bottleneck — the harder part is continuously collecting real-world data, building accurate world models, and letting them guide experiments.
The reasoning goes: while larger models have improved reasoning quality, automating experiments and decision-making hits a ceiling without a 'model of the world' that reflects physical constraints and causal relationships. Whether research automation heads in this direction can be read as a sober counterargument to scale supremacism.
The 'Manifesto Wars' Over Open Weights
On the ideological front, the conflict over open-weight models has entered a phase being called the 'manifesto wars.' The question of how much of a model to release has reportedly already moved beyond a technical choice and is becoming a clash of values.
One side champions full release in the name of reproducibility and external validation; the other argues for phased or restricted releases on grounds of safety and competitiveness. Both claim to act for the 'healthy development of AI,' and positions have hardened in a way that resembles political platforms. This conflict looks likely to cast a shadow over future regulation and corporate strategy as well.
Chinese AI Labs Are Making Four Different Bets, and Hermes Bets on Openness
Also eye-catching is an on-the-ground report that Chinese AI labs are currently playing 'four different strategies' at once. Companies differ in whether they emphasize scale, efficiency, specialization, or openness — suggesting that a monolithic view of a single 'China model' fails to capture the reality.
On the open side, NousResearch continues to roll out its agent-oriented 'Hermes' lineup. Community interest remains high, and practical agent models that run on workstation-class GPUs are said to be arriving one after another. While giant models dominate the headlines, the steady maturation of open agents that run on your own hardware is another important current.
Summary
The irony of a company flush with profits being the one to question the AI industry's trustworthiness, Microsoft's pragmatic push to turn a screen recording into an agent, the world-model debate pushing back against scale, and an open-weights argument taking on the color of an ideological battle. The developments of August 4 tell a story of diverse vectors advancing at once — one that can't be told in terms of giant models alone.