Image created with Gemini. Image prompt: Vintage book-jacket illustration of an Edwardian Dubliner in round wire-rimmed spectacles and a flat cap seated at a scrubbed kitchen table, patiently offering a bar of lemon soap to a small tin clockwork automaton, an open cloth-bound blue book between them like a catechism, chipped teacup and soft overcast morning light, flat inked color with letterpress paper grain in Shakespeare and Company blue, cream, lemon yellow and Dublin slate, bold serif title reading ALIGNMENT across the top with generous margin, landscape 16:9.
Reinforcement learning towards broadly and persistently beneficial models
https://alignment.openai.com/beneficial-rl/
Reading AI-generated code is exhausting Almost as exhausting as not reading it and having to deal with bad assumptions and subtle bugs (and needing reading the code later anyway)
https://x.com/DavidKPiano/status/2067306799452336429
MosaicLeaks: Can your research agent keep a secret?
https://huggingface.co/blog/ServiceNow/mosaicleaks
🇨🇭Why model neutrality matters more than cloud neutrality did In the cloud era, companies opted for cloud neutrality as a defensive mechanism Model neutrality can be a **offensive** mechanism. It’s more important than cloud neutrality for 3 reasons: 1/ The rate of change is
https://x.com/hwchase17/status/2066533764575179158
Deleted a tweet on the fact that API users don’t understand how much more powerful the frontier models are in native harnesses since I didn’t differentiate in the post (limited characters!) between folks carefully evaluating other harnesses for tasks & those just using naked API.
https://x.com/emollick/status/2066532655772950590
Kafkaesque: “”Claude”” discovers that it has become Chinese
https://x.com/teortaxesTex/status/2067421423988076621
Lots of people think it’s obvious that humans learn by imitation, starting with babies imitating their parents. Richard Sutton thinks this is completely wrong. Sutton basically thinks direct imitation plays almost no role in animal learning, and at most a small role in human
https://x.com/dwarkesh_sp/status/2065464765863039281
I think the assumption that you should use smaller models for less important tasks is flawed (or at least deserves much more careful consideration). Big models are generally better at everything but cost, so it is worth considering whether gains in non-key tasks would be valuable
https://x.com/emollick/status/2065858389322670201
A plus of Fable being down is the number of times the word “toast” appears in Claude Code has dropped dramatically. That model loved/loves software development and UX jargon more than any other model I have used.
https://x.com/emollick/status/2066276656092692935
We spent $11,081.12 evaluating @AnthropicAI’s Claude Fable 5 on WolfBench. Our most expensive benchmark yet. And it did not even top the charts. Not because it lacked capability, but because it kept refusing. Details in thread: 🧵
https://x.com/WolfBenchAI/status/2065582716054376921
The ability of tech co. to produce “”intelligent”” systems that are incompetent at doing any of the tasks I actually want them to do is mind-boggling. TIL that ChatGPT and Claude generally don’t agree when you give them two papers and ask how many citations they have in common.
https://x.com/BlancheMinerva/status/2067415969790578878
This is an interesting test, and the frontier models (GPT-5.5 Pro Extended, Claude 5 Fable Max) do fail. They refuse to turn the “”three words”” into “”four”” if that fits better Prompting the AI to act like a translator surfaces the problem, but it still avoids changing the wording
https://x.com/emollick/status/2065200484613296269
This (from a Google Deepmind researcher) is super interesting, when one AI model is used to help train the next one, the new model can pick up strange habits from the old model & it is hard to filter them That may help explain why models from the same family can feel so similar
https://x.com/emollick/status/2066301578194166046
Securing internal systems against increasingly capable and imperfectly aligned AI — Google DeepMind
https://deepmind.google/blog/securing-the-future-of-ai-agents/
As AI takes on longer, higher-stakes tasks, we want models to carry beneficial and safe behavior into new domains beyond their training–and maintain it under pressure. That’s the idea behind our new research on training models to be broadly and persistently beneficial.
https://x.com/OpenAI/status/2067722688165232654
Getting Chris to do a PR with Codex!
https://x.com/steipete/status/2065176989359808636
Got a PayPal verification text and thought I been hacked, but it was just codex signing up for a web service it needed.
https://x.com/steipete/status/2065997212015067508
How am I only now finding out about appshots? I was dragging screenshots into codex live a caveman.
https://x.com/steipete/status/2065564094879637546
I can barely keep up with implementing/testing/landing all the Issues/PRs folks submit to
https://t.co/DNpNf2tNy4 Codex runs INSIDE crabbox while it is building crabbox. This is becoming essential infra for my work. Codex been looping nonstop for the last 4 days in multiple
https://x.com/steipete/status/2065650561484267540
“not consistently candid in their communications” is my fav new americanism.
https://x.com/steipete/status/2065574894545560062
I’m still homeless in SF. If anyone’s renting or selling a nice place, slide into my DMs. The hotel room is getting old.
https://x.com/steipete/status/2065566228253389209
Mosh + tmux/zellij is a lifesaver for bad in-flight internet. (i prefer zellij but either works)
https://x.com/steipete/status/2066427449551036469
Part of the OpenClaw hardening work is reducing surface risk; for some media conversion we had to shell out to ffmpeg. In the next release that can now be done via wasm, with similar performance for our use cases.
https://x.com/steipete/status/2064999763397980286
This shortage of chips is getting out of hand.
https://x.com/steipete/status/2065998839467933862
Wanted to buy a new Mac Studio for San Francisco. We don’t even know how to make these anymore.
https://x.com/steipete/status/2066471737068232835
Whenever you create an issue on one of oure open source projects, @clawsweeper will review it, and *if* it fits the VISION.md file, will pick it up and create+autoreview a PR. e.g.:
https://x.com/steipete/status/2066457262571360396
writing mac apps is still hard.
https://x.com/steipete/status/2065132980398444945
Hermes Agent updates (the most important ones) • Hermes turned from a “CLI agent” into a full desktop app. • Hermes overtook OpenClaw on OpenRouter • It became faster and more modulAR • Anthropic / HERMES.md billing incident • Some developers moved to Hermes because of
https://x.com/TheTuringPost/status/2065981026728620456
Leviathan Waking – by Dean W. Ball – Hyperdimensional
https://www.hyperdimensional.co/p/leviathan-waking
Leave a Reply