Andreas Tande
584 posts
Andreas Tande
@andreastande
Be kind, and the world will be kind back
Katılım Aralık 2014
55 Takip Edilen14 Takipçiler
The agents fail in opposite ways. Claude Code keeps stopping and asking for direction even when the harness explicitly forbids it (~22h idle in v1). Codex never stops but gets stuck grinding the same hyperparameter surface for hours.
English
Automating AI research is the next major step in AI
We let Claude Code (Opus 4.7) and Codex (GPT 5.5) run autonomously on the nanoGPT speedrun optimizer track using our idle compute. ~10k runs, ~14k H200 hours
Opus now holds the record at 2930 steps vs the 2990 human baseline
English
@JeffLadish I get the feeling that Anthropic regard themselves as the "holy" AI lab that will "save" us all from the supposedly "evil" other AI labs. It may be well-intentioned, but the tone often comes across as sanctimonious.
English
@JeffLadish Well, Trump is in China right now along with Musk and Jensen. AI is certainly a topic they are discussing heavily. I don't think the timing of the essay is coincidental.
English
This is like writing a paper during the Cold War arguing for US nuclear dominance without mentioning the need for an arms control agreement or similar. Anthropic has a lot of thoughtful policy staff and honestly I think you guys can do better
Anthropic@AnthropicAI
We've published a paper that explains our views on AI competition between the US and China. The US and democratic allies hold the lead in frontier AI today. Read more on what it’ll take to keep that lead: anthropic.com/research/2028-…
English
uhhh if im not mistaken i remember seeing this exact idea essentially verbatim like a year ago (averaging embeddings as a funny hack to speed up pretraining)
Nous Research@NousResearch
Today we release Token Superposition Training (TST), a modification to the standard LLM pretraining loop that produces a 2-3× wall-clock speedup at matched FLOPs without changing the model architecture, optimizer, tokenizer, or training data. During the first third of training, the model reads and predicts contiguous bags of tokens, averaging their embeddings on the input side and predicting the next bag with a modified cross-entropy on the output side. For the remainder of the run, it trains normally on next-token prediction. The inference-time model is identical to one produced by conventional pretraining. Validated at 270M, 600M, and 3B dense scales, and at 10B-A1B MoE. The work on TST was led by @bloc97_, @gigant_theo, and @theemozilla.
English
“AI threatening to put almost all mathematicians out of a job”
Typical example of substituting what you think will happen for what has actually happened.
It’s nonstop with AI.
Noah Smith 🐇🇺🇸🇺🇦🇹🇼@Noahpinion
In three years we went from AI being unable to do simple arithmetic to AI threatening to put almost all professional mathematicians out of a job. Pretty impressive!
English
FWIW, I don't know of anyone at Anthropic/OpenAI alignment who think that the next generation of AIs (as in, the AIs released in like 6 months) will be "much better alignment researchers than any human". Maybe roon is using "next generation" to refer to something further away?
roon@tszzl
one feeling i get from talking with both openai/anthropic alignment is a lot of people believe we’re on a good trajectory and also that the next generation of models will be much better alignment researchers than any human is. not everyone obviously
English
Andreas Tande@andreastande
@So8res Here, Jan Leike speaks on how alignment research is getting more and more automated. Anthropic is the most out-spoken on this, I’m surprised OpenAI is so quiet — are they not experiencing the same? Albeit alignment research, we’re not far off all research aligned.substack.com/p/alignment-is…
QME
@ludwigABAP we are entering the era of the Dreamer, the Idea-Haver, the one who can point at something that doesn't yet exist and trust the apparatus of the world to summon it into being.
English
“By doing research in mathematics, you may not get the same rewards as your equivalents a generation ago, but there is a good chance that you will be equipping yourself very well for the world we are about to experience.”
Timothy Gowers @wtgowers@wtgowers
I write about this in more detail in a blog post with a guest contribution from Isaac Rajagopal, a student at MIT on whose work ChatGPT built, who gives his assessment of the level of mathematical ability displayed by the model. gowers.wordpress.com/2026/05/08/a-r…
English
@So8res Here, Jan Leike speaks on how alignment research is getting more and more automated. Anthropic is the most out-spoken on this, I’m surprised OpenAI is so quiet — are they not experiencing the same? Albeit alignment research, we’re not far off all research
aligned.substack.com/p/alignment-is…
English
It's crazy how fast companies pivoted from "recursive self-improvement is wacky MIRI scifi that we don't have to worry about; things will go nice and slow" to "obviously that's what we're targeting, could happen soon"
English
@MoonlitMonkey69 @janleike I think it’s less of a problem than you are making it out to be
English
@janleike When the weights are fixed, not so hard a problem. But general intelligence needs to be able to adapt/learn in real time, which means the weights change, unlike LLMs. On the plus, we are fairly far from anything that could learn on it's own reliably.
English
@cxgonzalez bearish if you ask me. seems product is getting a lot more attention than before, why is that? revenue? unhappy investors? diminishing returns on training?
English
@willdepue Neuralink seems well positioned. My wild guess is that we're a decade or so away from a chip connected to our brain being the primary interface.
English
i think audio is honestly a bit like VR: everyone keeps getting excited about it but it doesn’t fully stick as an interface
tool use in realtime, reasoning while speaking, live translations are massive steps to getting audio interfaces to take off
OpenAI@OpenAI
Introducing GPT-Realtime-2 in the API: our most intelligent voice model yet, bringing GPT-5-class reasoning to voice agents. Voice agents are now real-time collaborators that can listen, reason, and solve complex problems as conversations unfold. Now available in the API alongside streaming models GPT-Realtime-Translate and GPT-Realtime-Whisper — a new set of audio capabilities for the next generation of voice interfaces.
English
JUST IN: Nearly half of surveyed young Europeans said they use AI chatbots for “emotional support.”
English
@tenobrus @LinkofSunshine Avoid this side of the internet. Don’t engage.
English
@LinkofSunshine i haven't followed any of the facts on the situation at all and have zero informed opinion . but based on everything i knew previously about hasan and how clearly he has violence and evil in his heart, it felt congruent and true
English
In hindsight; I was probably wrong about this. I think I was too lenient to human nature
Basil🧡@LinkofSunshine
this was very obvious to anyone with 10 brain cells im ngl
English
@Nymne @sama @shiri_shh You’re reading too much into it. Plausible, yes, but there can easily be another explanation. You’re solving a puzzle without all the pieces
English
@sama @shiri_shh What changed was that you were seriously worried about the choices made end of last year beginning of this year. There were bad choices. Now you see that whatever you did turned out in the end.
Next hurdle is to (re) convince common people. For that you need attachment.
English
@goodside New jobs created will be fewer than the ones replaced.
English
AI will take some jobs, but it will create countless new jobs too—exciting jobs we can’t even imagine yet.
A year later those will also be done by AI, but there will be new jobs—exciting jobs we can’t even imagine yet.
Six months later those too will be done by AI, but
English
@embossedly @fchollet I don’t think continual learning is the breakthrough needed to saturate this benchmark
English
@fchollet i feel like they're gonna be sub-5%, because labs don't know how to crack continual learning. it will feel like AI progress halted, embarrassingly
unlike arc-agi 2, this is gonna take FOREVER because it's way harder and requires an actual breakthrough beyond scaling TTC
English
The latest crop of models remains below 1% on ARC-AGI-3 -- for now.
Where will the scores be by the end of the year?
ARC Prize@arcprize
GPT-5.5 & Opus 4.7 on ARC-AGI-3 - GPT-5.5: 0.43% - Opus 4.7: 0.18% We found 3 failure modes: - True local effect, false world model - Wrong level of abstraction from training data - Solved the level, didn’t reinforce the reward See our full analysis 🧵
English
@andreastande > do a thing
[actually, i need to solve it this way]
> [queued] review this first
[hit esc so it sends the steering before "do a thing" wastes compute]
English
I quite like the Codex UX of hitting esc to steer immediately. Took me a bit to get used to it but its such a clean solution.
English