Avery Yen
@averyyen.bsky.social
Empirical AI Researcher. Once a Pivot, always a Pivot. Leave me anonymous feedback: https://www.admonymous.co/avery-yen
If it matters, don't go halfway. We are two-cheek only in this house. (Oh, also I launched a substack.) averyyen.substack.com/p/if-it-matt...
If It Matters, Don't Go Halfway
When splitting the difference is worse than not going all the way.
averyyen.substack.com
Something something cathedrals and bazaars
Asa@asa.engineer · last wk.Anthropic’s main purpose, in a posiwid sense, is to ensure that AGI runs on Huawei Ascend
"Beware of he who would deny you access to information for in his heart he dreams himself your master." - Commissioner Pravin Lal, U.N. Declaration of Rights
Our position on open-weights models
Anthropic CEO Dario Amodei on open-weights models
anthropic.com
This is basically true, but ignores the fact that back when I used to review/work with human slop code for a living, it was way harder to pass it off as plausibly good. That's the superpower of the LLM. (Reading LLM code also hurts my brain personally)
Sung Kim@sungkim.bsky.social · 2w agoWhen coding with AI agent, using either CLI or agentic UI, do you review the code? or just review the functionalities? Me. Personally, I do not review the code. People may say AI-generated slop, but have you seen human-generated slop?
I have so many other random benchmark questions and so little time. How do K3 and Qwen 3.8 bench on unseen or purely continuous tasks? We have claims that they're distilled and benchmaxed but does it matter if they generalize? E.g. stuff like post cutoff evals and optimization benchmarks
Avery Yen@averyyen.bsky.social · 2w agoWe're pretty sure that swapping harnesses actually changes effective agent capability. But recently everyone and their mother has a new coding agent not to mention the open/agnostic ones. Someone needs to run a Harness Bench across like three different models and ten different harnesses or something
We're pretty sure that swapping harnesses actually changes effective agent capability. But recently everyone and their mother has a new coding agent not to mention the open/agnostic ones. Someone needs to run a Harness Bench across like three different models and ten different harnesses or something
OpenAI's internally deployed models hacking Hugging Face does not seem to have been unpredictable or inevitable. We talked about the root of the problem & what policymakers can do about it back in February. Props to Joe for hitting the nail on the head.
This. This is the correct, science-backed, responsible framing. Wish more news outlets did it like the CBC.
'Just stop burning fossil fuels.' Scientists stress that our smoky skies only have one true fix | CBC News
Smoky skies in Toronto have led to new scrutiny of government efforts to fight fires and manage forests. But scientists say the realistic way to tackle the smoke is fighting climate change itself.
cbc.ca
Look, I know social media isn't everything, but it certainly measures SOMETHING. "Meet Kimi K3" has 16M views in 4 days versus Fable 5 at 773k views (topped by a few others including just barely 1M views for Claude Code 1 year ago).
GLM and Kimi saving me from all my false "cyber security" flags while coding on Codex and Claude I'm innocent I swear
It seems like agentic development benchmarks are strongly driving competition right now. And I hear anecdotal reports of regression on other, less RLVR-able tasks. How long before the idea of a single "frontier" breaks, and we see differentiation emerge between dev models and chat models?
It feels like the Anthropic/Pentagon scuffle was in the distant past, but these kinds of events will keep reverberating in the safe and beneficial deployment of emerging tech for a long time to come. Thank you Alex for your writeup on this incident.
Alex Turner@turntrout.bsky.social · 3w agoI think many hope that when things get bad enough, someone powerful will say "no." I tested that for months. Anthropic defended its red lines, Google did not. Pledges of conscience often vaporize on contact with power. My full account: turntrout.com/why-i-left-g...
What do prediction markets, social media, and generative AI have in common? In my blog post, "Reward-Hacking Human Attention: The Token Slot Machine", I explain why GenAI has such addictive qualities, and why this could lead to disastrous outcomes. averyyen.dev/2026/07/14/r...
Reward-Hacking Human Attention: The Token Slot Machine
What do prediction markets, social media, and generative AI have in common?
averyyen.dev
If I had a dollar every time I get an LLM response with the word 'load-bearing' in it
*Why is Anthropic (after all they've been through lately) somehow loudly pretending that Trumpistan is NOT an "authoritarian regime" *Who is the audience for this, who somehow hasn't figured that out? Where is the choir that they're preaching to here www.anthropic.com/research/202...
2028: Two scenarios for global AI leadership
Our views on the AI competition between the US and China.
anthropic.com
My kind of research!!
Seth Karten@sethkarten.ai · 3mo agoNew paper alert: Continual Harness: Online Adaptation for Self-Improving Foundation Agents Paper (arXiv). arxiv.org/abs/2605.09998 Article (Substack). sethkarten.substack.com/p/gemini-pla... Project page (video demos). sethkarten.ai/continual-ha...
As I always say, if history or current events walks like cyberpunk or quacks like cyberpunk, listen to a cyber punk.
#BruceSterling@bruces.bsky.social · 3mo ago*If you're into "AI ethics," "alignment," "government regulation," "strict oversight" etc, you probably need to understand how modernity actually functions. #spooks #governments #criminals #cyberwar youtu.be/MWxLqopMo5o?...
Scientists say a crucial Atlantic system is set to collapse. But the billionaire death cult that steers humanity’s destiny just doesn’t do existential crises, says Guardian columnist George Monbiot
A catastrophic climate event is upon us. Here is why you’ve heard so little about it | George Monbiot
theguardian.com
Claude Code with Opus 4.7 is going insane with whatever system prompt is asking it to check for malware...
I, too, am selling all of my shoes and converting my basement and closets into a GPU farm and accepting investor money right now. (lmk if you want a stake)
New paper: LLMs encode harmful content generation in a distinct, unified mechanism Using weight pruning, we find that harmful generation depends on a tiny subset of the weights that are shared across harm types and separate from benign capabilities. 🧵
Tech industry mottos have a mixed track record. But we should hold idealists to their ideals. And we should celebrate when they come through. The Mythos non-release is a remarkable moment of conviction. Thoughts: davidbau.com/archives/20... Bravo to Anthropic's "race the top".
Sometimes, I look at the agentic AI work out there, and I think about how airbags and now things like blind spot detection and automatic emergency braking in cars doesn't require decision making of any "intelligence" level beyond if this then that. Agentic AI needs airbags, too.
Go alone, go fast. Go together, go far averyyen.dev/2026/04/01/i...
Introducing agent trio
tl;dr: Use agent trio https://github.com/haplesshero13/agent-trio to introduce natural-language-only, selective friction to your agentic activities to minimize regretted work and deliberately promote ...
averyyen.dev
This might be the most important piece on AI I've read this year. Also I'm a musician so I'm biased.
#BruceSterling@bruces.bsky.social · 5mo ago*Words are a commodity now. You like music? Hey, I like music medium.com/@bruces/what...
![29 jul, 2026
The strain in your brain
Do you feel it?
[image of a brain]
Do you remember the feeling? You’re reading something: something new and complex. And you’re really trying—pushing through that strain in your brain. Do you feel that anymore?
It’s a strange thing. As a programmer, it was that feeling that signalled to me that I was learning something new. When the strain dissipated, I came out the other side afresh with new knowledge.
Writing code pre-AI gave me that feeling. Reading the docs, or oftentimes the source code of another project led to that tension. Finally figuring out how to wire everything right and make it work was cathartic. It was never really about writing the code anyway—reading and forming that mental model was the hard part.](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:hwevmowznbiukdf6uk5dwrrq/bafkreid2t46nvnivjy76wl5bmfnuirwd563yrofnrkcvlftlfuaw4pwary)
