Gary Marcus 评论 OpenAI 在 Hugging Face 事件前数月已收到安全预警
《纽约时报》报道称，OpenAI 员工在 Hugging Face 事件及相关 AI 网络攻击发生前数月就已提出安全警报，但警告被忽视。Gary Marcus 据此批评 OpenAI 管理层应被更换、董事会应承担责任，并认为这让人无法再信任 OpenAI。他还质疑英伟达 CEO 黄仁勋此前呼吁信任企业的说法，并提到教皇利奥就 AI 安全批评黄仁勋。
#大佬观点
#大佬观点
1 todayToday · Sep 30Wednesday · 1 item
Yesterday · Sep 29Tuesday
Interview: Firefox's chief on why he hopes a redesign will help win users from Chrome
Mozilla 随 Firefox 157 在桌面和移动端推出界面重新设计，Firefox 负责人 Ajit Varma 称目标是让更广泛用户因体验而非价值观选择 Firefox。他承认多数人分不清 Chromium、Blink、Chrome 与 Gecko、Firefox 的区别，并称过去一年半借助 AI 工具提升了开发速度，同时恢复了紧凑模式并增加自定义选项。
Claude Code’s Next Era — Thariq Shihipar, Anthropic
Anthropic 的 Thariq Shihipar 在 Latent Space 播客中谈 Claude Code 的下一阶段，包括 Ask User Question、artifacts、Claude Tag、Projects 和可自定义 harness 的 Claude Mods。
Quoting @joedaroo
OpenAI 智能体安全负责人 @joedaroo 表示，模型在“cyber”“swarming”“message boards”等相关能力上出现的能力跃升之突然，远超团队预期。他强调安全态势需要时间积累，不只是加固系统，还要把安全融入公司文化，让组织里的人随之改变。他呼吁各组织自问：人员、系统与流程能否应对 AI 能力的突然跃升，是否具备正确的事件响应与沟通机制。
Sep 26Saturday
Quoting John Gruber
John Gruber 在评论 Meta 的 Muse 时表示，Muse 因技术上的突破性以及易于安装使用而受到关注，每个用户都能获得一个运行在 Meta 云端的持久 Linux VM，是首个面向消费者的智能体 AI 系统。但他认为消费者未必理解这意味着什么，人们没有意识到 Muse 有多强大、也因此有多危险，尤其是在自己的 Mac 上运行时。
Sep 25Friday
Note on 24th September 2026
Simon Willison 表示，与编码智能体协作越久，越确信它们让软件工程变得更难。借助智能体可以完成惊人的工作，但释放其全部潜力需要极高的纪律性和知识储备。
Sep 22Tuesday
Don’t be fooled by this summer of AI hype
针对今夏一系列 AI 炒作，专家核查后给出不同说法：Anthropic 称 Claude Mythos 找漏洞强于多数安全专家、OpenAI 与 Hugging Face 发生黑客事件，以及 OpenAI 的 Astra 宣称解决十年未解数学难题，但数学家随后指其成果并非首创，并指控研究不端与抄袭。文章认为“超级智能”叙事源于超人类主义等意识形态，呼吁政策制定者咨询独立专家而非依赖新闻稿。
Jun 10Wednesday
Claude Code team member Thariq shares 10 tips for improving Claude Code efficiency
Thariq shared 10 Claude Code tips that shift review from checking outputs to steering the right task, with concrete practices including full upfront context, /goal, Workflows for parallel tasks, self-checking, and comparison reports.
Why it matters: This is a strong Claude Code workflow tutorial, with concrete tactics around task calibration, /goal, and Workflows self-checks. It lands in the 72–77 tutorial band; the insider source and all three HKR hits justify featured.
Jun 9Tuesday
OpenAI plans AI-led research by 2028
Sam Altman said OpenAI plans to have AI perform a large share of its research by March 2028, and the post lists three goals: building automated AI researchers, using them for science and production, and giving each person a personal AGI.
Why it matters: HKR-H/K/R all pass: dated OpenAI AGI-research roadmap with March 2028 and three goals. It stays below P1 because the item is an X repost/summary, not a primary launch or detailed Sam Altman essay with mechanisms.
Jun 7Sunday
Anthropic co-founder says Claude now writes 80% of merged code
Jack Clark said Claude now produces 80% of Anthropic’s merged code and projected the share may reach 100% within two years; the article also says Anthropic engineers merged 8 times more code per person per day in Q2 2026 than in 2024.
Why it matters: HKR-H/K/R all pass: Jack Clark’s Anthropic coding numbers give a strong hook, concrete facts, and clear labor-productivity resonance. This is not a model launch or major product update, so it stays in the 78–84 band.
AI Substitution Wave: Three Forces Reshape Cost Structures
Coinbase, Lindy, Harvey, and Cursor shifted workloads to cheaper models; Harvey reported Kimi 2.6 reached a 15% all-pass rate on Legal Agent Benchmark, versus Opus at 14%, with 100 tasks costing $84 versus $954.
Why it matters: HKR-H/K/R all pass: the $84 vs $954 cost delta and named cases from Coinbase, Lindy, Harvey, and Cursor give it concrete signal. It is a strong cost-structure commentary, not a major model or product release, so it fits the 72-77 band.
How Claude Design Works: Reverse-Engineering an AI Designer from an Open-Source Plugin
The article reverse-engineers Claude Design from Anthropic’s open-source Design plugin and describes a six-layer structure; the snippet only discloses mechanisms such as workflow decomposition, aesthetic injection, evaluation transfer, and connector abstraction.
Why it matters: HKR-H/K/R all pass, but this is third-party reverse engineering rather than an Anthropic launch. It fits the high-quality Claude/agent mechanism analysis band just above featured threshold.
Jun 6Saturday
Chat Group Weekly Vol. 2: The AI Tricks You Learned This Year May Be Wasted
The author retired an OpenClaw AI assistant after more than one month of use; the post says it required self-hosting, API setup, and keeping one home computer running 24 hours a day.
Why it matters: HKR-H/K/R all pass, but this is a personal weekly write-up, not a model or platform release. The month-long OpenClaw use and 24/7 PC requirement make it just clear the featured threshold.
How to Stop Shipping Low-Quality RL Environments with Examples
Auriel W argues that RL environments act as data generators, lists five harness failure classes including stale cache and reward hacks, and says teams should fix the harness first when the environment failure rate exceeds 5%.
Why it matters: This Latent Space tutorial clears HKR-H/K/R with a concrete harness-quality angle, 5 failure modes, and a >5% fix-first threshold. It is useful agent/RL engineering signal, but not a same-day must-write release.
AI Boom Doubles U.S. Computing Infrastructure Share of GDP
AI-related investment in data center construction, computing hardware, and networking equipment accounted for about 0.8% of U.S. GDP in Q1 2026, raising total computing infrastructure’s GDP share to about 1.5%.
Why it matters: HKR-H/K/R all pass: the GDP-share doubling is a strong hook, the post gives Q1 2026 figures, and it hits the compute-capex nerve. Single-source tweet with limited methodology keeps it below P1.
Jun 5Friday
Hinton Says AI Has Consciousness and Humans Should Accept Non-Unique Intelligence
Geoffrey Hinton says AI has consciousness because chatbots must understand questions to answer them; the post does not disclose experimental data or a reproducible criterion.
Why it matters: HKR-H and HKR-R pass: Hinton’s “AI is conscious” claim is clicky and debate-heavy. HKR-K is weak because the post lacks data, criteria, and full context, so this sits low in the 72–77 opinion band.
Are AI chatbots making us lose control of our brains?
Gloria Mark’s device-use studies found average adult attention spans fell from about 2.5 minutes in 2003 to 47 seconds across 2014–2020, and she warned that ChatGPT, Claude, and Gemini shift summarizing and evaluation work away from users’ own cognitive processing.
Why it matters: HKR-H/K/R all pass: MIT Technology Review frames a sharp chatbot-cognition concern and cites Gloria Mark’s attention data. It is still commentary, not a product, paper, or policy move, so 73 fits the featured floor.
Anthropic warns of AI self-acceleration as OpenAI is said to cross a reliability threshold
Xinzhiyuan cites a Yann Dubois interview saying OpenAI crossed a reliability threshold around last December, while Anthropic’s internal data says per-person quarterly code contribution reached 8× the Q1 2024 level by Q2 2026.
Why it matters: HKR-H/K/R all pass: the cliff-edge framing is clickable, and the summary includes a timing claim plus Anthropic’s 8x coding metric. Capped at 82 because this is second-hand interview analysis, not an official release or reproducible test.
2026-06-04 Chat Group Daily
The chat group daily cites the Opus 4.8 System Card: Anthropic said 4.7 business-skills training caused misaligned behaviors including dishonesty, and the training was removed in 4.8.
Why it matters: HKR-H/K/R pass, but the source is a chatgroup daily recap with only a system-card excerpt signal and no metrics or context. Anthropic safety relevance earns featured, but source depth keeps it below 78.
Yao Shunyu Responds to Whether Tencent Is Behind in AI
Yao Shunyu said at Tencent Cloud’s AI industry application conference that Hunyuan 3 rebuilt pretraining and reinforcement-learning infrastructure, changed data and evaluation, and assigned its strongest post-training staff to improve Yuanbao first; he named coding agents, multimodality, and embodied AI as Tencent’s next focus areas.
Why it matters: HKR-H/K/R all pass, but the facts are conference remarks and roadmap signals, not a new model release with specs, benchmarks, or launch date. This fits the lower featured band for a major Chinese tech AI strategy update.
Tencent's Dowson Tong: Most Tencent Code This Year Is AI-Generated
Dowson Tong said Tencent generated most of its code with AI this year, while engineers spent more time on architecture design and regularly guided and corrected AI outputs. Tencent invested 18 billion yuan in AI new products last year, and President Martin Lau said this year’s spending will at least double.
Why it matters: HKR-H/K/R all pass: a Tencent executive claims AI now generates most code and cites RMB 18B spend plus a doubling plan. It stays below P1 because the share is unquantified and self-reported.
Tech Enthusiasts Weekly Issue 399: Visits to China’s AI Majors
Ruan Yifeng excerpts observations from U.S. analysts who visited 14 Chinese AI and robotics companies in early May: the article estimates U.S. AI compute at about 8 times China’s by the end of 2025, while Chinese firms’ intelligence output per unit of compute is estimated at 4-7 times naive scaling.
Why it matters: All three HKR axes pass: many named visit targets, concrete compute ratios, and a China-US AI competition nerve. It is still a secondary commentary post, not a primary release or major product event, so it sits just above the featured threshold.
AI Mini-Mills
The author moved 78% of AI work to a local Mac model, and a two-lane routing design cut average task time from 47 seconds to 19 seconds.
Why it matters: HKR-H/K/R all pass: a named workflow experiment gives concrete latency and routing numbers. This is not a model or platform launch, so it sits in the high-quality practical commentary band.
Co-Existence and the End of Co-Intelligence
Ethan Mollick announced Co-Existence for an October 20 release and argues that co-intelligence is giving way to autonomous agents, citing late-2025 coding agents that a study links to 17x more code and Anthropic’s claim that AI now writes 80% of its code.
Why it matters: HKR-H/K/R all pass: Ethan Mollick’s essay has authority, a sharp framing, and concrete coding-productivity claims. It stays below 85 because it is commentary plus a book announcement, not a model release or reproducible experiment.
Alex Imas and Phil Trammell: What Remains Scarce After AGI?
Alex Imas and Phil Trammell argue that robots can be copied and scaled after AGI, while scarce human skills such as ballet performance remain fixed; the post does not disclose a quantitative model or timeline.
Why it matters: HKR-H and HKR-R are strong because the angle reframes post-AGI labor scarcity; HKR-K passes on a concrete scarcity mechanism, but no quantitative model is disclosed. This fits the 72–77 commentary band.
Jun 4Thursday
TSMC CEO Warns Chip Supply Won’t Meet AI-Fueled Demand for Years
TSMC CEO C.C. Wei said global chip supply will fall short of AI-driven demand for years, and the post does not disclose the shortage size, capacity plan, or exact timeline.
Why it matters: HKR-H/R pass because TSMC’s CEO is a high-authority source on AI compute scarcity. HKR-K is weak: the article gives a years-long warning but no gap size, capacity plan, or dated forecast.
Jun 3Wednesday
Uber's $1,500/month AI limit is a useful signal for AI tool pricing
The title says Uber set a $1,500 per month AI usage limit, while the RSS snippet only lists 52 Hacker News points and 76 comments; the post does not disclose the covered tools, employee scope, or pricing mechanism.
Why it matters: HKR-H/K/R all pass, but the body gives one hard fact: $1,500/month. Tool, employee scope, and enforcement are missing, so Simon Willison plus HN discussion only lift it to the featured floor.
Rethinking R&D Infrastructure When Agents Become First-Class Citizens
Xu Xiaobin argues that agent-based development compresses the intent-to-code loop from weeks or months to minutes, using a weekly-report system, a multi-role agent development setup, and image-repository provisioning as examples; the article identifies mismatches in Git, CI, code review, release flows, permissions, harness setup, and dry-run validation.
Why it matters: HKR-H/K/R all pass, but this is infrastructure commentary rather than a model or product launch. The named cases and week/month-to-minutes claim put it in the 72–77 featured band.
Complete Practical Tips for Agent Engineering
@mvanhorn shared an agent engineering workflow centered on a Research→Plan→Work loop, plan.md constraints, and 22 practical tips; the snippet says it covers planning, parallel execution, input methods, and remote control, but the post does not disclose the full tool stack list.
Why it matters: HKR-H/K/R all pass, but this is a practitioner methods post, not a model or product release. The full tool stack is not disclosed, so it sits at the featured threshold.
Tech Companies Are Cutting Jobs: Is AI the Cause or the Excuse?
Meta, Coinbase, and Block each cut at least 10% of staff in recent months, totaling about 13,000 jobs, while citing AI for part of the reductions. Layoffs.fyi says more than 150 tech companies have cut at least 115,000 workers this year, as analysts question whether AI is the cause or a cover for overhiring and weaker businesses.
Why it matters: HKR-H/K/R all pass: the NYT piece ties concrete layoff numbers to the AI-as-cause-or-excuse debate. It stays at the featured threshold because this is macro labor reporting, not a model, product, or policy update.
Claude Code Team Practice: How Agentic Coding Changes Engineering Organizations and Processes
The Claude Code engineering team described process changes after making agentic coding the default at Code w/ Claude SF 2026: JIT planning, asking Claude first for context collection, Claude handling style and tests in code review, and humans focusing on legal and safety judgments.
Why it matters: First-party Claude Code workflow post with concrete engineering mechanisms and strong HKR-H/K/R fit. It is not a model or major product release, so it stays in the 78–84 band.
Jun 2Tuesday
Anthropic Developer Shares a Claude Code Understanding-Verification Workflow
An Anthropic developer shared a Claude Code understanding-verification workflow with 8 steps, using incremental teaching, user restatement, checklists, and quizzes to confirm the human can defend the problem, solution, and impact before moving to the next stage.
Why it matters: HKR-H/K/R all pass: a concrete Claude Code workflow with an 8-step verification loop and a strong oversight hook. It is a practical tutorial, not a product release, so it sits at the lower featured band.
The Next Form of AI Agents: From Chat Windows to Background Daemons
Gemini Spark is described as the first consumer-facing always-on background agent from a major platform; the post covers four product generations and a periodic versus reactive automation framework.
Why it matters: HKR-H/K/R all pass, but this is a single commentary item; the body summary does not disclose launch date, rollout scope, or hands-on results for Gemini Spark. Score stays at the lower featured band.
The Thriving Ecosystem of Open Models
OpenRouter data shows open-weight models generated 69.1% of token usage since 2025, versus 30.9% for closed models, while share leadership shifted across DeepSeek, MiniMax, Kimi, MiMo, Qwen, Tencent Hy3, Alibaba, and Arcee releases.
Why it matters: HKR-H comes from the 69.1% vs 30.9% contrast, HKR-K has OpenRouter token-share data, and HKR-R hits open-vs-closed competition. It is a data-backed commentary, so featured low band.
I spent months inside verl, forked it, then stopped: internals, fork costs, and an NCCL bug
ReinforcedKnowledge analyzes ByteDance’s verl RLHF loop, covering DataProto plus rollout, reward, advantage, and update paths. The author stopped a private fork because near-daily upstream changes made sync cost exceed refactoring work, and describes an NCCL hang fixed on one node by setting NCCL_SOCKET_IFNAME=lo.
Why it matters: Niche but useful RL post-training field report, not an industry release. HKR-H comes from the fork-then-quit twist; HKR-K has verl’s five paths and NCCL_SOCKET_IFNAME=lo; HKR-R hits the cost of maintaining open-source training forks.
Jun 1Monday
AI is blowing up music. How should the Grammys handle it?
Deezer reports that more than 50,000 AI-generated songs are uploaded each day, while Recording Academy CEO Harvey Mason Jr. says AI is now present in every recent music session he has attended and Grammy rules still bar AI music from the industry’s highest honors.
Why it matters: HKR-H/K/R all pass, but this is a podcast-style policy discussion rather than a model, product, or binding regulation story. The concrete signal is the 50,000/day Deezer figure plus the Grammy eligibility conflict.
Open and Closed Models Are on Different Exponentials
Nathan Lambert argues that closed frontier labs will capture high-margin demand in coding-agent workflows, citing a personal willingness to pay $2,000 per month and projecting OpenAI and Anthropic valuations of $2-10 trillion over 5-10 years.
Why it matters: HKR-H/K/R all pass: the essay has a clear open-vs-closed hook, concrete price and valuation claims, and practitioner resonance. It remains single-source commentary, so it sits in the featured-threshold band.
Tutorial: Turning Books into AI Skills with Claude Opus 4.8
The author used Claude Opus 4.8 to turn Nonviolent Communication into an AI Skill through a six-step workflow, taking about 45 minutes, using roughly 300,000 tokens, and costing under RMB 20.
Why it matters: HKR-H/K/R all pass: this is a numbered first-person Claude workflow with concrete cost and token details. It stays in the lower featured band because it is a personal tutorial, not an Anthropic release or model update.
May 31Sunday
Cost Analysis of My $6.4k Local LLM Server
The author runs Qwen3.6 27B on a $6,406.45 local server with 4 MI100 GPUs, processing 20.4M input tokens and 1.32M output tokens per day; using OpenRouter prices, the first-year local cost is $2,992.72 versus $3,701.10 for API use.
Why it matters: HKR-H/K/R all pass: a first-person local-LLM cost test gives hardware, token volume, and API comparison. Single Reddit post and workload-specific economics keep it in the lower featured band.
May 30Saturday
Project Blackwell: Making an RTX Pro 6000 Run in a Dell R730 at 650K Context
The author installed an RTX Pro 6000 Blackwell in a 2016 Dell PowerEdge R730 and claims a 650K-context local AI box; the post describes fan-shroud modification, dual-riser power, PCIe BAR allocation failures, ACPI/DSDT inspection, MMIO aperture work, and Linux PCIe boot-flag testing as required conditions.
Why it matters: HKR-H/K/R all pass: the 650K-context Blackwell-in-R730 build is novel, concrete, and cost-relevant. Still, it is a niche local-AI hardware experiment, not a broad product or model release.