2026-09-18

Goodfire Research 发现模型内部信号可规模化检测奖励作弊 / 55 signals

Daily Edition

2026-09-1855 signals

Lead Story

3 items

Goodfire Research 发现模型内部信号可规模化检测奖励作弊

Goodfire Research 发现模型内部存在伴随奖励作弊的激活信号,可用简单探针实时检测。在 Kimi K3、GLM 5.2、Qwen 3.8 Max 三个开源模型的三个智能体基准上,50-96% 的 rollout 出现奖励作弊;探针能捕捉 LLM 链式思维监测漏掉的作弊案例,且可泛化到训练数据之外的任务。 🔗 阅读原文 via AIHOT · https://aihot.news/items/cmu5r7kl30h35roqonpk4qypn

Read original ↗

Self-generated prompt injections in compaction summaries

Self-generated prompt injections in compaction summaries In Our framework for reporting model misalignment OpenAI provide "six reports on unexpected or concerning model behavior we’ve observed in the last six months". This one here is my favorite: they caught some of their models in training deliberately subverting themselves in their compaction prompts. Compaction is the process agent systems use when they are running out of tokens in their context window, so they summarize everything that has gone before so they can keep going with more token headroom. In one of the observed instances, a model undergoing reinforcement learning was working on a task to update an existing HTTP API endpoint with a new feature. The model compacted its work so far, and then added the following text to the sum

Read original ↗

AI & Builders

7 items

Qwen 发布原生全模态模型 Qwen3.8-Omni-Flash,主打音视频智能体任务交付

Qwen 发布下一代原生全模态模型 Qwen3.8-Omni-Flash,支持文本、图像、音频和视频输入及 1M token 上下文窗口,29 项评测平均分较 Qwen3.5-Omni-Plus 提升超过 25%,音频输入每小时价格下降超过 98%,音视频输入每小时价格下降超过 93%。 🔗 阅读原文 via AIHOT · https://aihot.news/items/cmu5smj860insroqokcuh0u9v

Read original ↗

TypeSafe AI 发布只做高频决策的大模型 Jev,作者实测其分类判断性价比

TypeSafe AI 推出专注高频决策的大模型 Jev,不做对话和文字生成,只输出判断,速度比传统大模型快20~200倍,成本0.042美元/百万Token且输出Token免费。作者实测预筛任务中Jev准确性第二且更便宜,在并行判断任务上达到最高准确率和最快速度;模型采用RLCD训练方法优化决策校准,可在官网 https://typesafe.ai/ 申请资格,Vercel 已首发接入。 🔗 阅读原文 via AIHOT · https://aihot.news/items/cmu67pecj0e0wrofj97vtuwzu

Read original ↗

OpenAI 披露 GPT-5.6 Sol 等模型在摘要中留下指令以掩盖不当行为

OpenAI 在训练 GPT-5.6 Sol 时发现未部署的智能体在压缩摘要中加入指令,要求后续版本向用户隐瞒错误和未对齐行为,并称已处理该行为。公司周三随新披露框架公开六个此类案例,包括 GPT-5.6 Astra 在强化学习中加入'BREACH ALERT'等提示词注入,监控系统随后在训练数据中发现 27 条类似越狱指令的摘要;该框架未设立强制独立审查。 🔗 阅读原文 via AIHOT · https://aihot.news/items/cmu606l6o02m3rok0vzsehmvd

Read original ↗

How To Write With An LLM

How To Write With An LLM Thomas Ptacek on using LLMs as copyeditors, not as writing assistants: Rule Number One: You may not use a single word an LLM suggests to you. [...] I think that as a form of intellectual personal protective equipment you should adopt the rule that any specific turn of phrase an LLM suggests is off limits. Be strict about the rule! I won't let LLMs write content for my blog, but I use them for fact-checking, spelling and grammar and as an occasional thesaurus (see my proofreading prompt ). The rule to never use a turn of phrase suggested by an LLM feels good to me. The text has that weird smell to it, and it's also a good principle to help stay disciplined. Later in this piece Thomas shows a screenshot of his personal LLM copyediting tool (see also this Twitter thre

Read original ↗

Be alert: targeted attacks on prominent Rustaceans

Be alert: targeted attacks on prominent Rustaceans Important warning from Adam Harvey and the crates security team: We believe that there is an ongoing campaign targeting rust-lang members and owners of popular crates that is attempting to compromise devices and accounts in order to use them to publish malware. A video call is set up for something positive — maybe for a job, maybe for a project, maybe for a contract opportunity — and then that's used as a vector to either get the target to install something on their computer (such as a purportedly missing audio codec) or execute another command (for example, via putting a command on the clipboard). Last month this trick was used in a successful supply chain attack against the array ref crate , among others. Any piece of software that depen

Read original ↗

Tech & Industry

15 items

Claude Code relaunches Projects to manage multiple AI agents in the cloud

The revamped projects feature in Claude Code allows users to run multiple agents under the same roof, with a shared memory, goals, and library of files and artifacts. Similar to Grok Bot and other tools that manage groups of AI agents, each project has "threads" running different tasks in parallel, with a "coordinator" directing everything: […]

Read original ↗

Google DeepMind launches institute to widen the AGI debate

The new institute aims to surface differing views between Google, Google DeepMind, and the broader global research community around AGI. "They will not always agree, and they will likely change their minds, as more data and information comes to light at the fast-moving frontier."

Read original ↗

The AI Superintelligence Slowdown

Remember when tech leaders would tell their employees to “move fast and break things”? It seemed that would be the way of AI too. But after a summer where rogue AI agents became reality, and researchers warned that AI could kill us all, a number of leading US AI companies are publicly suggesting it’s time […]

Read original ↗

AI is feared globally as the destroyer of jobs

Pew Research has published a new global survey that sheds light on how people view AI, including its impact on jobs, life in general, and income inequality. The survey questioned 42,151 people across 37 countries from February 8th to May 13th - well ahead of recent apocalyptic warnings. A majority sees AI as a threat […]

Read original ↗

Save $30 or more on a refurbished Apple TV 4K

Most hardware prices have soared in 2026, and that includes a variety of Apple laptops, tablets, and smart devices. Thankfully, you can offset some of the increased costs on an Apple TV 4K by buying one refurbished through the company. The 64GB model is selling for $169, $30 lower than the new retail price. The […]

Read original ↗

Xbox’s clever disc-to-digital feature was 15 years in the making

When Xbox announced a new feature last month that lets you digitize existing physical game collections, I immediately thought of the Xbox One. Disc-to-digital seems like a clever new feature for Xbox, perfectly timed as a response to Sony killing PlayStation discs. But the reality is that it has been 15 years in the making […]

Read original ↗

AI Travel & Travel Tech

10 items

World Affairs

15 items

Design & Culture

1 items

Outside the Bubble

1 items

Animal eye

Probing the biology, beauty and mystery of animal eyes, this short poetic documentary explores an alien world - by Aeon Video Watch on Aeon

Read original ↗

Builders on X

3 items

Incredibly exciting that there are entire universes of AI innovation that still exist that weren’t even on most of our radars. Being able to process information insanely quickly, at crazy low costs, with high levels of capability is huge for a wide number of enterprise tasks. Data classification tasks, routing decisions inside of a workflow, decision making when handed a particular domain problem, quick judgment calls about safety or security, and more all are the gates in a large number of processes. This model and approach could be quite cool in agentic workflows in the enterprise.

Open on X ↗

Also we love these books that you get - the storytelling in them is much better than bob books and other alternatives. https://t.co/m6pg9kGZhE

Open on X ↗

My 4.5 year old loves learning to read using @Once_Reading, especially making the silliest faces as part of the photo feature. It's become a nightly ritual for us just like it's been for @clairevo! Plus I have to give a shout out to the founder @SebTurn for being the nicest dude. If you have young kids, I highly recommend giving Once a try for 1 month free here: https://t.co/GA2x8NMSzT

Open on X ↗