fchollet: Model Instruction Execution Decoupled from Decision Making
Keras founder notes rapid improvement in instruction-following but stagnation in autonomous decision-making beyond instructions.
There is an interesting disconnect between the ability of models to successfully execute precise instructions — improving incredibly fast — and their ability to make sound decisions when faced with something not covered by the instructions — which has been stagnating for a while. This gap between obedience and judgment raises questions about the true nature of model progress.
AI Labs Avoid Commodification via Vertical Integration
Analysis shows AI companies are moving up the stack, using enterprise deployment and switching costs to achieve lasting profitability.
Current model inference businesses face commodification: models are converging, switching costs are low, and price competition is squeezing margins. Labs are responding by going vertical — embedding into enterprise workflows, building proprietary deployment pipelines, and creating switching costs that lock in customers. The strategy mirrors historical patterns where infrastructure providers were commodified while software layers captured value. The risk is reduced competition and enterprise lock-in.
The term AGI has aged very poorly. AI is far better than human intelligence for most economically-relevant tasks. But that doesn't make AIs better than humanity.
— Guillermo Rauch, Vercel CEO
Kimi Business Membership Launches for Enterprise Orders
Supports 5-seat minimum, corporate bank transfers, self-service invoicing, and includes all Kimi Allegretto plan features with enterprise-grade support.
Claude Code Weekly Limit Boosted 50% Until August
Claude Code maintains a 50% weekly limit increase for Pro, Max, Team, and seat-based Enterprise users through August 19.
Claude Fable 5 Permanently Built into Max Plan at 50% Cap
Anthropic announces Fable 5 will be permanently included in Max and Team Premium plans with a 50% usage cap, likely driven by pressure from OpenAI and Kimi K3.
Kimi K3 Ranks Third on VoxelBench, Just 100 Elo Behind Fable
Kimi K3 jumps from K2.6's 28th place to third, trailing Claude Fable 5 by approximately 100 Elo points on the VoxelBench leaderboard.
Kimi K3 Tops SpreadsheetBench 2, Surpassing Claude
An open-weight model now outperforms all proprietary models on AfterQuery's SpreadsheetBench 2, with Kimi K3 ranking first ahead of Claude Fable 5.
Grok 4.5 Appears on Cognition's FrontierCode Leaderboard
Cognition launches the FrontierCode leaderboard, a new benchmark for code generation, with Grok 4.5 showing strong early performance.
Grok TTS Scores 94 on Humanness Index, Just 6 Below Humans
Grok's text-to-speech model takes the top spot on The Humanness Index with a score of 94 for naturalness, only six points below the human baseline.
Grok 4.5 Uses Only 1.5T Parameters, 3× Cheaper Than Kimi K3
Grok 4.5 operates with 1.5 trillion parameters versus Kimi K3's 2.8T, while costing three times less per task — demonstrating intelligence efficiency.
Anthropic Quietly Reduces Fable 5 Effective Cap by 25%
From July 20, the Fable 5 usage cap is reduced by 25% versus the promotional period, effectively halved from the 150% limit users enjoyed since May.
Mind Lab Open-Sources 2M Token Long-Context RL with Only 8 GPUs
Mind Lab's 2M-token long-context reinforcement learning project is now open source. Remarkably, it required only 8 GPUs compared to the thousands-of-GPU solutions previously thought necessary, dramatically lowering the barrier for long-context training.
DeepSeek Accused of Mass Scraping Proprietary Model Outputs
Allegations claim DeepSeek may be systematically collecting outputs from multiple proprietary models in a large-scale operation, sparking ethical and technical debate across the AI community about data provenance and fair competition.
Codex Computer Use: Auto-Installs Blender, Creates 3D Animation
Ethan Mollick demonstrates Codex's computer control capability: "I've never used Blender, download and install it using your computer control and then use it to make an otter in 3D." Only one click was needed to grant installation permission.
Codex Team Already Generates Most Code with AI, Humans Act as PMs
Codex team members reveal that most daily code is now generated by Codex itself. Developers increasingly act as product managers and QA, defining features and verifying results rather than writing implementation details.
Claude Design Prototypes Replace Traditional Product Docs
Interactive high-fidelity prototypes from Claude Design are more intuitive than traditional documents. AI implementations achieve over 90% fidelity from these interactive mockups with simulated data, making them a superior specification medium.
Anthropic Mass Account Bans Spark User Outcry
Many long-term Claude accounts, some years old, are being banned in a new enforcement wave. Users report that even newly purchased accounts are not spared, generating significant community frustration.
Chinese AI Labs' Capital Efficiency Far Exceeds US Counterparts
Nathan Lambert observes that Chinese AI labs are far more capital efficient. In a world where intelligence is proportional to effective capital for compute, data, and talent, this efficiency may be their greatest strategic advantage.
Open Weights Won't Cause AI Collapse, Compute Providers Capture Value
Ethan Mollick argues that even if open-weight models eventually dominate, compute remains the bottleneck. Value would shift to compute providers rather than model labs, and AI itself would not become useless — the economics simply redistribute.
US Labs Underestimate the Power and Vitality of Open Research
Graham Neubig agrees that US labs are undervaluing open research, citing Richard Hamming's classic talk "You and Your Research" as an essay that rings true at both micro and macro levels about the unpredictable power of openness.
Kimi K3 came closer to Fable than Anthropic could do with Sonnet 5. The point about American legal scruples is underrated — down with the red tape!
— @teortaxesTex, on Kimi K3's distillation breakthrough
Kimi K3 Built on KDA Architecture, K3.1 Expected to Surpass Fable
KDA is an extension of GDN with vector-based decay, proven at scale. K3.1 is predicted to surpass both Sol and Fable 5 when enough SuperPoDs are available.
DeepSeek Suspected of Routing API Traffic Through Claude
Accusations claim DeepSeek may be using its own API as a proxy to access Claude — if confirmed, the implications would be devastating for trust in the ecosystem.
OpenAI Build Week Hosts 32 Global Community Events
OpenAI Build Week holds 32 community events worldwide this weekend across Bangkok, Bengaluru, and more, focused on building, exploring, and applying AI tools.
ACT-2 Preview: First Robot Model Unifying Generalization and Reliability
A single fine-tuning checkpoint adapts ACT-2 to tasks across diverse embodiments — a robotics foundation model combining broad generalization with high reliability.
OpenAI Faction Claims Open Weights Equal Deceleration and Communism
OpenAI proponents argue open weights slash frontier margins and reduce capex for AGI progress, dismissing the ecosystem as decelerationist and harmful to innovation.
OpenAI Strategy Team Warns Kimi About Open-Sourcing Powerful Models
Kimi reveals receiving an open letter from OpenAI's strategy department cautioning about safety risks of releasing high-intelligence open-weight models to the public.
Who Pays for AGI? Beyond Oligopoly, New Answers Needed
Ilya Sutskever testified, "If there is no funding, there is no big computer." The question of who funds AGI if no oligopoly exists cannot be answered with quips.
AI Capability Gains Do Not Automatically Yield Economic or Social Benefits
A powerful case study shows increasing AI capabilities require deliberate deployment and governance to translate into real societal improvements.