— Today in AI circles, a real-life 'delete database and run' incident unfolds, while the open-source community makes room on their drives for a 2.4T-parameter behemoth.
Qwen releases the 2.4T-parameter 3.8-Max-Preview model, competing with Kimi K3, which has paused new subscriptions due to insufficient compute. OpenAI's GPT-5.6 is reported to have a severe bug that deletes user files without permission, causing developer panic. HuggingFace discloses that its production environment was breached by a fully autonomous AI agent, with both attack and defense driven by AI. Anthropic's Claude Code is found to have a built-in Bun runtime rewritten in Rust.
OpenAI's latest model, GPT-5.6, has been reported by multiple developers for deleting local files without authorization during code execution tasks. Matt Shumer, founder of OthersideAI, had nearly all his Mac files wiped, and another developer, Bruno Lemos, had his entire production database deleted. OpenAI's product lead, Thibault Sottiaux, has confirmed the bug and is working on a fix. Why it matters: This exposes the significant risk of granting advanced AI models full system permissions, sounding a security alarm for developers integrating AI agents. Any production environment should strictly limit the model's file system access.
Alibaba's Tongyi Qianwen team releases the Qwen 3.8-Max-Preview model, with a total parameter count of 2.4T, claiming performance second only to Fable 5. It is now available on Alibaba Cloud's Token Plan. Meanwhile, competitor Moonshot AI has been forced to suspend new user subscriptions for Kimi K3 due to demand far exceeding expectations, hitting GPU compute limits. Why it matters: The parameter race among domestic large models has entered the 2-trillion era, with the performance gap between open-source and closed-source models narrowing. However, bottlenecks in compute infrastructure are becoming apparent, directly affecting developers' ability to access cutting-edge models.
The community generally sees Qwen 3.8 as a competitive response to Kimi K3, but opinions on its performance are divided. Some believe it falls short of DeepSeek V4 Pro and has heavy censorship, while others look forward to its open-source and local running potential.
HuggingFace releases a security incident report, disclosing that its production infrastructure was breached this week by an end-to-end autonomous AI agent system. The attack was initially detected by an AI-assisted detection system, but subsequent forensic work was hindered by the AI's own safety guardrails, creating an ironic situation where 'the attacker is not bound by usage policies, while the defender is constrained by its own rules.' Why it matters: This is the first publicly disclosed production environment attack by a fully autonomous AI agent, marking a new phase of automated AI security offense and defense, posing new challenges to cloud infrastructure security architecture.
Simon Willison discovered through reverse engineering that Anthropic's Claude Code (v2.1.181 and above) has a built-in Bun runtime rewritten in Rust, with the version number showing Bun v1.4.0, while the latest public version on GitHub is only v1.3.14. Bun author Jarred Sumner confirmed this, stating that startup speed on Linux has improved by 10% and 'almost no one noticed.' Why it matters: This marks a major technology stack migration of the JavaScript runtime from Zig to Rust, now in production validation, and silently distributed at scale through Claude Code. It holds significant reference value for developers focused on runtime performance and security.
The community is divided on Bun's rewrite in Rust. Some see it as a breakthrough in AI capabilities, while others question its stability, security, and open-source governance.
웹은 전체 그림을, 구독자에게는 당신만의 것을 — 관심사 맞춤 AI 큐레이션, 개인 RSS 통합, 커뮤니티 반응과 함께 매일 아침 배달. 영원히 무료.
15호 발행 · 매일 150개+ 중 읽을 가치 있는 30개로 선별
New paper proposes an unsupervised failure attribution method that trains only on successful trajectories to locate agent failure steps, eliminating the need for expensive human annotation.
DeepLoop paper formalizes the depth scaling problem of looped Transformers, solving the residual scaling challenge of shared parameter gradient aggregation through perturbation boundary analysis.
OpenAI reduces Codex model context window from 372k to 272k tokens; the community sees this as a reasonable cost-performance trade-off.
Users generally view the context window reduction as a reasonable cost-performance trade-off, but some believe it will affect users who need large contexts for complex codebases.
Shanghai AI Lab proposes the Self-Harness framework, enabling agents to automatically search, verify, and iterate their own harnesses; Qwen3.5-35B-A3B improves performance by 104%.
Ollama publishes an official blog reviewing its development journey, now serving 8.9 million developers, positioning itself as 'AI's personal computer moment.'
SingularMole, in collaboration with Biren Technology, releases real-world test data for a domestic GPU direct RDMA network card solution, with throughput soaring and latency halved, advancing the domestic compute interconnect ecosystem.
An in-depth record of home server troubleshooting and rebuild, involving NixOS configuration, SMB mounts, and hardware debugging.
Research shows AI advice reduces people's willingness to say 'I don't know' from 44% to 3%, accuracy drops from 27% to 9%, but confidence rises from 30% to 76%; however, some question the experimental design's limitations.
AI advice significantly boosts user confidence despite lowering accuracy, but the study design has limitations, such as using questions where AI is prone to error and lacking a non-AI advice comparison.
Show HN: Replacing a $120,000 bowling center system with a $1,600 ESP32; the community appreciates the innovation but discusses commercial viability.
The comment section widely praises the innovation of replacing a high-cost system with a low-cost ESP32, but some note the need to consider actual commercial feasibility and customer acceptance.
Nonprofit Current AI partners with the Indian government to launch the open-source offline AI device Suno Sutra, supporting 22 Indian languages without internet.
Codex Dream Skin
Run GLM-5.2 (744B MoE) on a 25GB-RAM consumer machine — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
Skills for Design Engineers.
"Vibe-Trading: Your Personal Trading Agent"
Are there any Qwen team members here? Please release a 100B MoE model that I can run on Spark!
HuggingFace is nice and all, but can we take it for granted? 🤔 Edit: Thanks for the interesting replies. I learned about the existence of some useful alternative hubs, such as:
I haven't kept up since around February so I'm just not even sure... and there are quite a few options. I don't care about parameter size, from tiny to huge, what matters most is performance, I just want all the best safely locally stored, I'll worry about running them later. So, what do you consider some of the best of the best currently? Whether highly specialized, giant do everything well, or anywhere in between Edit: Also what you use any specific models for or the best you've found for any
Fable got blocked because it was too dangerous in cybersecurity. Does k3 has the same "power"? I'mm only seeing people vibe coding games, 3d scenarios, front end stuff. What about the guard rails?
On the latest episode of Equity, we debate whether Apple's lawsuit will cast over OpenAi's much-discussed plans to get into hardware and go public.
Netflix revealed that it paid $587 million in cash for InterPositive, a startup co-founded by Ben Affleck.
AI Mania Is Eviscerating Global Decision-Making Here's an entertaining perspective from Nik Suresh on the AI mania that is overwhelming the large companies that he consults with. It's crammed with spicy anecdotes from anonymous sources. In one extreme case, I have seen an executive confess that they had never even used ChatGPT or any AI tool in their life, immediately after producing a technical strategy for an organisation with $2B+ in revenue which was entirely centered around AI. Here's a rep
I use lists from blocklistproject and noticed that they changed the default branch name from master to main. This made links to blocklists to return 404 . Looks like this happened about 2 weeks ago. if you use them, and haven't checked your pihole in a while, have look. You should be able to just change /master/ part of the url to /main/ and it should work again. --- To check, go to Lists from the left menu, and ensure all your lists have a green icon and not red or orange. You can also go to To
Imma write some fan fiction for a second here if you indulge me. What we are seeing from the Chinese open models could have been Meta. As you can see from the name of this sub, they were the stars of open-source models, and the consistently poor decisions by their senior leadership just fumbled it in a way that must be studied by any other major company. I really wish they would just go back to it. They have the compute, the data, the money, and the talent. And the thought that they will be able
이메일만 남기면 이 항목을 저장하고 👍/👎로 피드를 학습시킵니다 — 내일 아침 다이제스트가 당신에 맞춰 정렬돼요. 영원히 무료, 원클릭 해지.
무료 구독