←
tokenbender15,366 posts
tokenbender avatar

tokenbender

@tokenbender

Founder - Deus eXperiments • Building apprentices that grow with you • avataRL, codeCherryPop, PIC, evolvedSeeker

◷ Joined July 2014▧ 2,141 Photos and videos
tokenbender avatar
tokenbender@tokenbender·Oct 13

Introducing Chantless: Chapter 1 — Apprentice A powerful tool makes you free, allows you to express yourself, and learns and grows with you. This is what we're building with Chantless. A technology that lets you shape and evolve models alongside you.

1:26
tokenbender avatar
tokenbender@tokenbender·6h

big RL doesn’t want you to know but this is exactly what training long horizon agents is like.

How is this even possible 🙂

1:34
Kyle🤖🚀🦭 avatar
⟳ tokenbender retweeted
Kyle🤖🚀🦭@KyleMorgenstein·19h

people kept asking me for “hot takes” at CoRL. I was sorry to disappoint, because what I came up with felt lukewarm: RL is the only thing that has ever worked at scale. exploration + adaptation. that’s it. everything else is a party trick. RL is the ultimate bitter lesson.

you may think “RL is not the right tool for this problem” and you would be wrong. it’s all RL. this is sick.

1:10
tokenbender avatar
tokenbender@tokenbender·10h

this is why i have lots of respect for people working on evals because it is a thankless job requiring lots of care. if it works fine, you are worried what you are missing and if it doesn't then you lie down, curl up and cry about loss of time and compute.

The evals are not ok New NeurIPS paper: Hidden Measurement Error in LLM Pipelines Distorts Annotation, Evaluation, and Benchmarking. Evals drive deployment, safety standards, and research. Yet we ignore noise from prompt, judge, agent harness etc, & so underestimate error in benchmarking & annotation, & it gets worse with more data. I'll get to what this means for CIs & how to fix, but first let's talk about how this makes benchmark hacking easier 🧵

Post attachment
james yu avatar
⟳ tokenbender retweeted
james yu@jamesjyu·16h

All this pangram controversy (ie. the latest Caine Prize) could be avoided by allowing writers to use whatever tool (AI included) they want and judging stories by their own merit This graph will only continue to grow and the literary world will tear itself apart

Post attachment
tuna🍣 avatar
⟳ tokenbender retweeted
tuna🍣@tunahorse21·20h

yeah sol and astra 6.2 dead on arrival got my hopes up, even reinstalled codex absolute gigafried RL nonsense speak models that make opus 5.5 look like genuine super intelligence

tokenbender avatar
tokenbender@tokenbender·21h

rip

Post attachment

Some of you have been asking very politely. Some of you have just been asking. GPT-6.2 Astra and Sol are now available. Super efficient and built to perform where it counts—the real world. openai.com/index/introducing…

1:59
tokenbender avatar
tokenbender@tokenbender·Nov 16

my unresearched crude framing of an idea is better than a researched but easy to produce information.

Post attachment
tokenbender avatar
tokenbender@tokenbender·Nov 16

there’s a “worship of effort” bias that all humans share. few years back, programs as well as essays used to be heavily researched from existing blogs/top links on googling something. what people really refuse to address is that in both of these cases, what really made the final result good was exactly what were the questions being asked or answered. what was being added at the edge of discoverable information and i don’t understand how that can’t be done with AI? i’m doing 20x or more experimentation than last year, and it takes more than 60% of my mental energy now in trying to process all the expweimental data into an article and make it “not ai” which is absurd. i understand the argument for cases where something is truly hard to absorb for the audience. most content produced by ai is vacuous and contains zero information while being hard to understand. so the real problems that need to be solved are: > understanding what’s truly added here that’s at the edge of discoverable information > lucidity of the written content needs to be maintained > maybe a disclosure at the top saying you converted your idea seeds or voice inputs to an essay. but the “worship of effort” bias isn’t going to go away and i suspect we’re going to need to build things to fool these detectors even though i find it all so tiresome.

CNN@CNNNov 15

The college admissions essay is in a very weird place. It's impossible to say just how many of them are now written by AI, but at this point, experts agree that it's a lot of them. And while colleges generally frown upon that practice, they're using AI to evaluate the essays, too. cnn.com/2026/11/15/tech/ai-s…

tokenbender avatar
tokenbender@tokenbender·Nov 16

i am done assessing carbon on my workflows. this is a great model and congrats to gemini team for this. easier to work with than openai or anthropic counterparts but also makes strange mistakes and its presence is irritating to both oai and ant models in agent teams for some reason.

*sigh* pays for ultra sub i need to sniff these carbon monoxide fumes now

wordgrammer avatar
⟳ tokenbender retweeted
wordgrammer@wordgrammer·Nov 15

Every great idea™️ was at one point an illegible, vague instinct in a great thinker’s mind. “I dunno but I think I’m on to something.” Then they figure it out, find the right words, and have crystal clear writing. And so people think “clear thinking” leads to good ideas.

Tensor-Slayer avatar
⟳ tokenbender retweeted
Tensor-Slayer@TensorSlay·Nov 14

Reprogramming dormant neurons : Qwen TTS Ever wondered why Qwen3-TTS sometimes breathes or laughs out of nowhere? Agent found neurons behind it. The details: Agent found 417 neurons to be specific (about 0.48% of the talker's MLP neurons) and 61 of them are dormant. Amplified together, they give the following effects effect: 1. Breathes as words are spoken 2. Hushed ASMR 3. Natural laughs as language is being spoken 4. Enhanced whispering 5. 2nd speaker appearing. Comparisons and neuron map repo in the comments below. Inspired by recent work on Lazarus by @DimitrisPapail , which reprogrammed dormant neurons.

0:15