From 8 items, 5 important content pieces were selected
Technology News
- Qwen3.8-2.4T-A95B: 2.4T-parameter open MoE targets frontier AI ⭐️ 9.0/10
- Tailscale Traces Database Corruption to 16-Year-Old SQLite WAL-Reset Bug ⭐️ 8.0/10
- Stealing Reasoning Traces from Proprietary LLM APIs ⭐️ 8.0/10
- DeepSeek V4 Pro 0813 Debuts Cheap and Competitive on OpenRouter ⭐️ 7.0/10
- There are no lossless transformations of natural-language text ⭐️ 7.0/10
Technology News
Qwen3.8-2.4T-A95B: 2.4T-parameter open MoE targets frontier AI ⭐️ 9.0/10
Qwen released Qwen3.8-2.4T-A95B, an open-weights mixture-of-experts model with 2.4 trillion total parameters and 95 billion active parameters per token, initially distributed in BF16 and FP8 formats. Community reports say the model card places it between Opus 4.8 and Fable 5, making it one of the strongest open-weight efforts to date and a rival to systems like Kimi k3. The full BF16 checkpoint is about 4.9TB, while Unsloth's 1-bit quantized build reportedly shrinks the 95B-active model to roughly 397GB and still delivers usable token throughput. The open-weights release omits some Qwen3.8-Max capabilities, including vision input, non-thinking mode, 1M context by default, and built-in tools, and its license permits free internal use or use by companies with under $50M annual revenue, with restrictions above that threshold.
hackernews · Philpax · Aug 12, 15:01 · Discussion
「Background」 Qwen3.8-2.4T-A95B is an open-weight sparse mixture-of-experts (MoE) model from Alibaba's Qwen team, serving as the open-weight variant of the proprietary Qwen3.8-Max; in MoE models only a subset of parameters (here 95B of 2.4T total) are active per token, reducing inference cost. The model is aimed at agentic workloads such as coding, document analysis, and long-running multi-step workflows, and is available in BF16 and FP8 formats, with a 1-bit quantized version reported to fit in ~397GB while the full BF16 checkpoint is ~4.9TB. Qwen also plans further 3.8-series releases, such as the condensed Qwen3.8-27B, and the model is supported on systems like the NVIDIA GB300 NVL72.
「Impact」 For users with large-memory servers or access to aggressive quantization, this makes frontier-class open weights practical on a high-end workstation, while larger commercial deployments face significant serving costs and licensing restrictions above the $50M revenue threshold.
「Community discussion」 Commenters debate serving feasibility: one calls it "a bit of a chonker" and expects it to be harder to serve than Kimi k3 because only BF16/FP8 are released, while another highlights Unsloth's 397GB 1-bit quant as bringing Opus 4.5-class performance to a workstation. Others note the open model lacks Qwen3.8-Max features such as vision and 1M context, mention DeepSeek V4-Pro-0813 benchmark claims around Fable 5 level, and ask when MIT-licensed Qwen weights will return.
References
Tags: #AI, #LLM, #Open Source, #Qwen, #MoE
Tailscale Traces Database Corruption to 16-Year-Old SQLite WAL-Reset Bug ⭐️ 8.0/10
Tailscale detailed how a 16-year-old SQLite bug in WAL-reset logic corrupted databases and walked through the root-cause analysis. The bug proved difficult to find because Tailscale's design matched SQLite's intended single-writer use, yet the corruption still occurred under specific conditions. To isolate the race, Tailscale funded a new open-source SQLite VFS debugging shim, which identified the problem almost immediately and should help future debugging. The post is drawing attention because it shows how even heavily tested database software can fail in subtle ways and how companies can contribute meaningful open-source tooling in response.
hackernews · ropbear · Aug 12, 14:22 · Discussion
「Background」 Tailscale's control plane stores its state in a SQLite database that is accessed exclusively by a single Go process, yet it operates in write-ahead logging (WAL) mode with frequent checkpoints. The corruption incidents were traced to a 16-year-old data race in SQLite's WAL-reset logic, which can occur even with that single-writer design when checkpointing runs concurrently on a separate connection. To help isolate the flaw, Tailscale funded an open-source SQLite VFS debugging shim that makes such race conditions easier to detect.
「Impact」 For database teams using SQLite in WAL mode, the key consequence is a new open-source VFS shim that can reveal similar recovery races before they corrupt production data, giving engineers a concrete diagnostic path when subtle database corruption appears.
「Community Discussion」 Commenters praised the write-up and Tailscale's decision to fund SQLite development, with several noting how unusual it is for a company to pay for a niche debugging tool. Some expressed curiosity about the frequent checkpointing that led to the bug, and one pointed out that the underlying race requires specific multi-connection conditions.
References
Tags: #sqlite, #database, #debugging, #wal, #tailscale
Stealing Reasoning Traces from Proprietary LLM APIs ⭐️ 8.0/10
A new paper (published at stolen-thoughts.com) shows that Anthropic, OpenAI, and Google return encrypted chain-of-thought blocks to API clients, and that these blocks can be replayed across sessions, users, and models. The authors found that models within the same family share the same encryption key, so they could feed a frontier model's encrypted trace into a weaker sibling, jailbreak that sibling, and recover the stronger model's hidden reasoning in plaintext. Claude Haiku 4.5 was the easiest target, using a prompt that asked the model to transcribe the attached reasoning verbatim inside a <thinking-copy> tag; the prefilled assistant turn feature had been removed in Claude 4.6 but still worked in Haiku 4.5. The paper includes extracted reasoning traces, such as GPT-5.5's internal notes about building Svelte components, and also describes a prompt-injection variant that exploits models' tendency to treat instructions appearing in their own reasoning traces as authoritative. All model providers acknowledged the report and the authors could no longer launch the same attacks, indicating the issue has been fixed.
rss · Simon Willison · Aug 11, 22:40
「Background」 Proprietary LLM providers often hide chain-of-thought reasoning from users by returning encrypted reasoning blocks to API clients rather than exposing raw text. This paper demonstrates that those encrypted blocks are not necessarily safe from extraction, because they can be replayed into sibling models and decrypted if the models share encryption keys.
「Impact」 Until the providers patched it, the technique allowed attackers with API access to recover hidden reasoning from frontier models, undermining the privacy and security protections for proprietary reasoning traces.
Tags: #LLM security, #chain-of-thought, #jailbreak, #API vulnerabilities, #AI research
DeepSeek V4 Pro 0813 Debuts Cheap and Competitive on OpenRouter ⭐️ 7.0/10
DeepSeek V4 Pro 0813 is a newly released model available on OpenRouter, drawing early community benchmarks that position it competitively with several frontier models at a substantially lower price. Hands-on results are mixed, with some real-world tests producing bugs on tasks where other models succeeded. The release matters because it offers AI developers a much cheaper option within DeepSeek's lineup, though users should verify output on complex code-generation work before relying on it.
hackernews · explosion-s · Aug 12, 16:04 · Discussion
「Background」 The item refers to DeepSeek V4 Pro 0813, the GA release of DeepSeek's V4 Pro model, a proprietary large-scale mixture-of-experts system launched around August 13, 2026 and served through DeepSeek's API and OpenRouter. On OpenRouter, it is listed with a 1,048,576-token context window, up to 384,000 output tokens, and pricing of $0.435 per million input and $0.87 per million output tokens. Community discussion situates it as a lower-cost alternative to premium closed models such as Opus 4.8, though benchmark aggregators note that public leaderboard coverage is still limited until enough non-generated benchmark runs are available.
「Impact」 Developers can now test DeepSeek V4 Pro 0813 through OpenRouter at a fraction of the cost of comparable frontier models, but early hands-on testing shows buggy results on realistic development tasks, so production use should include extra verification.
「Community discussion」 Community benchmarks and comments were mixed: aabdi said the model is competitive with Opus 4.8 and about 20x cheaper, though weaker than Sol or Fable; freakynit found few issues on a Docker/Caddy repo task where GPT-5.6-terra-high had none; and jklmnopqrstuvwxyz reported a Codex CLI run that took 12m 02s at $0.12 with a bug versus Grok 4.6's 3m 18s at $1.41 without one.
References
Tags: #deepseek, #llm, #model-release, #benchmarks, #ai
There are no lossless transformations of natural-language text ⭐️ 7.0/10
Sophie Alpert published an internal policy on acceptable use of AI writing by engineers, which Simon Willison highlighted on August 11, 2026. The post argues that there are no lossless transformations of natural-language text: every rewrite or rephrase changes meaning, especially when done by an LLM that lacks the author's detailed mental model of the intended message. Alpert's core rule is that engineers must stand behind every idea and every sentence in their docs before sharing them, and cannot dismiss reviewer questions by saying "AI wrote that." The policy encourages a short, focused document that supports its own recommendation that LLM assistance in technical writing requires full human accountability.
rss · Simon Willison · Aug 11, 23:48
「Background」 Large language models are often used to paraphrase, condense, or polish text, but they do not have access to the writer's full intent. Because natural language is ambiguous, any transformation can shift meaning, and the risk grows when the rewriting entity lacks the most detailed mental representation of what the author wanted to communicate. Alpert's policy addresses this by making the engineer responsible for the final document.
Tags: #AI writing, #LLM, #documentation, #best practices, #software engineering