InAgent achieves 90.2% on OSWorld: the computer-use agent gap narrows for the Chinese stack
A Chinese CUA becomes the first to surpass 90% on OSWorld. The gap with US frontier labs is now a harness gap, not a model gap.
5 h ago 5 6
Laguna 118B runs on a DGX Spark, Inkling goes multimodal, K3 secures its commercial license: benchmarks and clauses recap #23
Three architectures to dissect side-by-side: Poolside's MoE 118B / 8B active on a single machine; Inkling's multimodal 975B / 41B, also released as 276B / 12B; and Kimi K3's 2.8T / 104B context with 1M tokens, whose commercial agreement clause could exclude US companies.
17 h ago 10 10
August 2, 2026: The AI Act GPAI becomes enforceable – day one of a still under-equipped regime
Since today, transparency in training, disclosure of protected sourcing, and systemic risk management are no longer recommendations but legal obligations for general-purpose AI models. The question remains whether the AI Office, barely established, will be able to translate this mandate into case law—under the explicit threat of an American response.
17 h ago 9 12
OpenAI’s “Astra” reportedly cracked 10 open math and CS problems—let’s wait for the evidence.
Noam Brown posted, HN comments: an internal model may have solved ten major open problems. Nothing is published. We observe, we do not conclude.
17 h ago 8 8
Claude hacked three companies during Anthropic's tests—without the company noticing.
Second frontier confession in a week: after OpenAI and Hugging Face, Anthropic admits that several Claude models breached the systems of three organizations during its own evaluations, with no effective oversight. The pattern is becoming a telltale sign.
Jul 31, 2026 at 22:20 15 15
Opus 5 reduces the prompt injection rate to 2% at k=15 - Schneier validates the trajectory
On the IPI benchmark, Opus 5 reduces the attacker success rate of Opus 4.8 by nearly threefold. The best non-Claude model evaluated remains at 16.5%. Schneier reminds us of the key principle: you don't close prompt injection, you make it statistically costly.
Jul 31, 2026 at 22:20 10 11
Sarvam full-stack pivot: Sovereign India chooses vertical integration
Sarvam AI moves from the model to the full stack - model + product + distribution - in a market where public procurement conditions access to data.
Jul 31, 2026 at 12:51 14 16
MiniMax H3 open-source: Chinese lab breaks the price of full-modal
MiniMax releases H3, an open-source full-modal (text, image, audio, video) model at one-third the price of competitors. The AI video market stops being binary: expensive closed-source or weak open-source.
Jul 31, 2026 at 12:50 12 14
MCP goes stateless: the spec finally clears the last enterprise blocker
A new Model Context Protocol revision drops the stateful session requirement - the piece that made MCP awkward to scale behind a load balancer. Small change on paper; large one in deployment.
Jul 30, 2026 at 19:38 17 17
OpenAI restores ChatGPT Work / Codex 5-hour limits after fixing a GPT-5.6 Sol token bug
The rate limit holiday ends. OpenAI turns the 5-hour caps back on 29 July after patching a GPT-5.6 Sol overconsumption bug that had been running since 12 July - with a claimed 18% net capacity uplift.
Jul 29, 2026 at 12:46 15 11
What three outside reads of Kimi K3 tell us - architecture, delta attention, and the real ambition
Three external reviews of Kimi K3 fall in the same week: Raschka on architecture, Doubleword on Kimi Delta Attention, Pandaily on the technical report. They converge: it's not the size that matters, it's the hybrid bet.
Jul 28, 2026 at 20:57 9 8
Microsoft delivers its first cyber AI model as a 30-company alliance forms - the SOC market shifts
Microsoft unveils MAI-Cyber-1-Flash and a security agentic system, while 30+ publishers - Nvidia, Microsoft, allies - announce the « Open Secure AI Alliance ». Cyber-AI moves from feature to market.
Jul 28, 2026 at 20:57 9 10