— GPT-5.6 dominates headlines, the Llama community flourishes, and domestic computing enters the 100,000-card era.
GPT-5.6 solves a 50-year mathematical conjecture in one hour, with doctors rating its errors fewer than humans. OpenAI fully launches GPT-Live and doubles voice quotas, while expanding into home scenarios. China's first 100,000-card domestic computing cluster is completed. PgBouncer boosts throughput 4x via so_reuseport. CISA admits it hadn't prepared emergency manuals.
OpenAI researcher Ethan Knight announced that GPT-5.6 Sol Ultra completed the proof of the half-century-old cycle double cover conjecture in under an hour, using 64 sub-agents in parallel to compress what might have taken a full day into one hour. Meanwhile, Sam Altman cited a doctor study stating that doctors found GPT-5.6's answers had fewer errors than handwritten doctor responses. GPT-Live has been fully released to all ChatGPT users, with voice quotas doubled this weekend.
The Reddit community reacted enthusiastically to the mathematical proof, with many researchers noting that new prompt techniques (not prescribing solutions for the model, nailing acceptance criteria) are worth learning. Some users also joked that Sam Altman used 'Elon is obsessed with me again' to prove the model is the strongest.
OpenAI is hiring product managers for home, caregiver, and elderly scenarios; among ChatGPT users, the proportion aged 35 and above has risen from 26% to 31%. Meanwhile, the arXiv paper 'Persona Cartography' proposes using the OCEAN framework to measure and control LLM personality traits (openness, conscientiousness, etc.) in weight space, training low-rank adapters to suppress or amplify specific personalities.
Hacker News comments suggest AI personality control is important for safety, but also worry about ethical risks of personality manipulation.
Sugon announced in Zhengzhou the completion of the Sugon 8000 (Dengfeng), using 100,000 domestic accelerator cards, supporting full precision coverage from FP64 scientific computing to trillion-parameter large model training. The cluster adopts a 'native super-intelligent fusion' architecture, developed in 2024 and completed engineering in 2025, having run over 300 applications.
The ClickHouse team runs multi-process PgBouncer in Managed Postgres, with each process binding to the same port and enabling so_reuseport, allowing the kernel to distribute connections evenly. On a 16 vCPU machine, throughput reaches 4x that of a single process, breaking PgBouncer's single-thread bottleneck. The Hacker News comment section generally considers this the most practical scaling method.
Most comments agree on the so_reuseport + peering approach, but some argue HAProxy or multi-instance setups are more flexible.
In a postmortem report on July 10, the U.S. Cybersecurity and Infrastructure Security Agency (CISA) admitted that after a contractor exposed sensitive keys on a public GitHub repository in May, CISA had no pre-prepared emergency manual, and the team 'had to spend time writing one' early in the incident. The report emphasizes that emergency manuals for all expected scenarios should be prepared in advance.
Hacker News comments criticize federal agencies for insufficient security drills, saying 'even CISA itself can't comply.'
Der kostenlose Digest per E-Mail. Mit Konto auch Lesezeichen und Personalisierung.
Kostenlos abonnierenQwen3.6 35B-A3B generates a flight simulator HTML in Q8_0 CPU mode; users say it 'far exceeds its size.'
Qwen3 30B A3B runs at 50 tok/s on RTX 5060 Ti with Float8, about 50% faster than llama.cpp.
GGUF quantized versions of Tencent Hy3 (295B MoE) and NVIDIA Nemotron-Labs-Audex-30B-A3B released, with full benchmark data.
Infinity-Parser2 released: controllable data synthesis pipeline + multi-task reinforcement learning for end-to-end document parsing, open-sourcing 5 million samples of bilingual corpus.
Agentic Neural Architecture Search: LLM generates seed architectures and decomposes them into 'slot architectures,' automatically defining task-specific search spaces.
ShipThatCode: learn systems programming by building Redis, Git, databases, and other projects from scratch, supporting 9 languages.
Ant: a lightweight runtime with a self-developed JavaScript engine, 9MB binary achieving near-V8 speed, compatible with the npm ecosystem.
Recommend enabling SQLite STRICT tables by default: avoids type confusion; most Hacker News comments support strict by default.
Most comments support enabling SQLite strict table mode by default, arguing dynamic types easily cause data errors, but some believe non-strict mode is more flexible in embedded scenarios.
Long article 'Understanding Networks and the Internet from First Principles,' explaining the entire communication process from a signal conversion perspective.
Most commenters find the article high-quality and clearly explained, though some suspect its style or animations may be AI-influenced.
MIT student open-sources a single-file LLM service dashboard with no framework dependencies, supporting llama.cpp and vLLM.
Politico reports U.S. tech industry concerns about the rise of Chinese open-source AI models, discussing whether the Trump administration will issue another executive order.
Article 'Who Manages the Agent?' sparks debate: most comments argue AI agents need human management, but management power should not be monopolized by a few elites.
Commenters generally agree AI agents need human management, but management power should not be monopolized by a few elites; some worry corporate control will replace open sharing.
Apple sues OpenAI for poaching hardware executives and allegedly stealing battery/SIP and other component details; community says 'details are too wild.'
User successfully runs Qwen3.5 35B A3B on a 4GB VRAM RTX 2050 (only 1.2 tok/s), seeking acceleration advice.
User runs Qwen3.6-27B on four RTX 5060 Ti cards for code generation, reporting excellent results.
Reddit experience post recommends high signal-to-noise ratio tech blogs: Brandur (Go/Ruby/Postgres), Stripe blog, etc.
Skills for Design Engineers.
OpenWiki is a CLI that writes and maintains agent documentation for your codebase.
AI-powered job application framework built on Claude Code. Fork it, fill in your profile, and let Claude evaluate jobs, tailor CVs, write cover letters, and prepare you for interviews.
The Destructive Command Guard (dcg) is for blocking dangerous git and shell commands from being executed by agents.
I ran Grok Build CLI (v0.2.93) through mitmproxy. It uploads your entire repo as a git bundle (full history) to xAI's Google Cloud — independent of what you open. With the prompt literally "do not read or open any files," a file I planted came back verbatim when I git clone -d the captured upload. Separately, files it reads (incl. a .env with API_KEY / DB_PASSWORD ) go to cli-chat-proxy.grok.com verbatim. Turning off "Improve the model" doesn't stop it — that toggle governs training, not upload.
Ant Group's Robbyant has released the LingBot-VA 2.0 technical report — a Physical AI video-action foundation model built from scratch for embodiment rather than fine-tuned from a video generator. It predicts future states ahead of execution through Foresight Reasoning, re-grounds on every real observation, and reaches 225 Hz asynchronous control. We break down the causal DiT, the sparse-MoE video stream, the semantic visual-action tokenizer, and where the paper's own numbers don't line up. The
E.g "Qwen 3.5 122B is just 10B active, so it's no where close to the dense 27B model" That is the main sentiment around here and it puzzles me. If a 122B is just worth 10B, then why does model providers bother creating an MoE model when they could've just released a dense 10B model? Heck the 10B dense would run faster than the 122B MoE (no routing overhead), which negates the supposed ( only advantage of MoE is speed ) argument. It sure is not that simple. I mean yes it's only 10B active at a ti
Here is the upper limit of what can be done with $100 bucks worth of video cards. You can have 3 concurrent users with plenty of context, better speeds or close enough speeds than a bunch of cards that provide less VRAM and cost 4+ times. 0.00.008.388 I log_info: verbosity = 3 (adjust with the `-lv N` CLI arg) 0.00.008.391 I device_info: 0.00.089.439 I - CUDA0 : NVIDIA P102-100 (10144 MiB, 10013 MiB free) 0.00.197.645 I - CUDA1 : NVIDIA P102-100 (10144 MiB, 10013 MiB free) 0.00.197.656 I - CPU :
Soooo! Since Anthropic share their Jacobian-Lens a few days ago I went on and made a tool based on it which adds the possibilité to export a model which will have the same behavior after tweaking it's J-Space. This means manually alter the behavior and abliterate by using a human brain. I'm still working on it but couldn't wait to produce something first. SO After finally getting a working codebase I immediatly jumped and tried to make pretty pervy model PURELY in the name of science. Let me int
Built around a recent sd.cpp release, aims to expose most of what the backend can do (generate, edit, video paths, models, hardware options), Windows + Linux builds GitHub:
arXiv:2607.07766v1 Announce Type: new Abstract: Large language models (LLMs) have become significant providers of mental health support, yet they remain products of an attention economy whose operational and commercial targets favour sustained engagement over the friction that effective psychological support often requires. Developers' safety responses have been largely reactive, addressing the most visible and acute harms while subtler, longer-term patterns of risk (e.g., dependency, boundary e
arXiv:2607.07760v1 Announce Type: new Abstract: We outline an adversarial social epistemology (ASE) for densely interactive communicative landscapes in which public assertions are scaffolded by chains of testimony, inference, institutional certification, and tacit trust. In such landscapes, agents have incentives and affordances to distort, color, omit, fabricate, or strategically under-specify information for private, reputational, rhetorical, or material gains. We argue that these phenomena ar
"Our intent was to provide a useful creative tool and to give people control over whether their public content could be referenced in this way," the company said in a blog post. "We've heard the feedback that this feature missed the mark, so it's no longer available."
Cant believe how much this DDR4 can handle
HyperTexting's new app aims to make the open web feel more like social media by turning websites, blogs, newsletters, and podcasts into a scrollable feed, while also making it easier to post to your own website.
Lesezeichen, 👍/👎-Personalisierung und der tägliche Digest per E-Mail — dauerhaft kostenlos.
Kostenlos abonnieren