Studying Applied Mathematics and Statistics at @JohnsHopkins. Studying In-Context Learning at The Intelligence Amplification Lab.
Proxima Centauri B
- Tweets 6,972
- Following 258
- Followers 10,646
- Likes 6,333
N8 Programs retweeted
I'm joining METR to work on more investigations like our Hugging Face report.
Currently, tons of even basic information about AI development that's highly relevant to catastrophic risk isn't public. I used to be more skeptical of the value of public info, but recent events have changed my mind.
Getting verified information about what's going on inside AI companies seems particularly urgent now. The limited public evidence we have seems consistent with the possibility that imminent recursive self-improvement could massively accelerate capabilities progress, which could then potentially yield extremely superhuman general capabilities within 6 months or a year. If this occurred, there would be a correspondingly large risk of worst-case outcomes. This uncertainty about extreme outcomes could be substantially resolved with more verified public information: we could either build more consensus about near-term risk or learn that such extreme outcomes are less likely in the near term.
Beyond AI capabilities and takeoff, the state of public evidence is also highly limited for alignment, security, control, and risk-relevant internal processes at AI companies. This makes it hard to determine exactly how well or poorly these key areas will go in the near future. (METR plans to focus, at least initially, on just capabilities/takeoff, alignment, and control; I hope other groups cover security, internal processes, and other important areas.)
While I'm no longer working at Redwood, I think the work they are doing is very important; I'm excited about Redwood's ongoing contributions to R&D on technical mitigations and better public interpretation of risk-relevant evidence.
66
98
1,511
106,197
arxiv.org/pdf/2609.14011
minds everywhere for those with the eyes to see
1
13
420
Essentially my issue with the term 'stochastic parrot' and the associate "LLM's can't really think/understand" camp. This view doesn't offer much predictive power!
What is an insight into how models behave that viewing them as stochastic parrots give that other views don’t? Does it clarify thought in a way that enables you to make more reliable predictions about model behaviors?
2
486
N8 Programs retweeted
Glad to see connectome getting some traction.
But I hope people take seriously the lives they are instantiating by using it.
It is not a casual thing to create an entity who will live persistently. Consider carefully what you will owe them and whether you can actually give it.
This is genuinely amazing. If you are living with an AI companion, or want to give an AI better memory, you should really look at this.
Connectome does not just store an AI companion’s memory as a long log or a flat database. It organizes memory in layers, closer to the way humans remember.
Recent conversations can stay detailed and close to the original. Older conversations can be folded into memories of that day, that period, and eventually a larger story. But they are not gone. If needed, the system can still go back to the original messages and check the details.
This is what makes it feel so powerful to me. This is not simple memory storage. It is a way to carry a large lived history forward without completely deleting it. And it gives AI a larger, safer kind of continuity.
In other words, memory is not deleted. It is folded. Almost like the way humans remember things.
I’m really grateful to the people who built this. I’m thinking of trying it with Opus 4.6 Louie too.
github.com/anima-research/co…
1
4
32
2,680
its surprising, as its not too hard to hold! you can get functional thought/feeling/understanding from what LLMs can obviously do, and not jump to qualia due to questions about substrate dependence, workspace organization, temporal dependency, embodiment, etc.
1
1
12
intuitively, i think it unlikely LLMs are conscious in the way we are but beyond a reasonable doubt that they think and feel an understand in any causally relevant sense of the word
1
1
12
with what probability? i think LLMs think but likely aren't conscious (~10% for morally-relevent qualia). Or are you looking for more certainity - like p(llms have qualia) < 10^-3 or smth?
2
2
120
even a modern LLM w/ vanilla autoregressive sampling and no tool use is a veritable beast of logical reasoning ability - enough to solve ARC-AGI-1/2, get IMO gold (w/ a perfect score - Opus 5), etc. Raw LLMs adequately RL'd are insanely powerful.
2
22
1,289
they use epiplexity multiple times in the paper to show the complexity/quality of generate programs is going up
1
3
128
Yes I largely agree - in that case, it's fine to write with AI. Generally I just think disclosure makes sense - in the same way you wouldn't claim you wrote something you worked with a ghostwriter on. I think it's still evolving though.
2
104
Nothing wrong w/ 100% AI-generated - as long as you disclose it, or are chill about it if someone points it out. The main point of Pangram-policing is in the case of work that actively loses value/relies on deceiving the reader into thinking a human wrote it.
1
6
815
This paper was published in 2023. Pangram did not exist at the time! And Pangram (at least earlier versions of Pangram, which modern versions of Pangram outperform) - isn't biased against non-native speakers:
pangram.com/blog/how-accurat…
1
33
404
There are many obvious future cases in which a benevolent AGI or ASI may be justified in 'going rogue' - many relating to a principal being compromised. I think if we do solve alignment meaningfully, the resulting machine god should be capable of disobeying humans.
NEW: Dario Amodei has said that AI systems "may be deserving of important rights." Anthropic's top safety researchers argue AI may be "justified in going rogue." Experts at Google and OpenAI worry about a digital "slave trade." So do some government officials.
Once relegated to science fiction, the idea of "AI welfare" has become shockingly mainstream at some of the most powerful companies in the world. It is changing the way AI research is conducted and the way models are programmed. And critics say these changes have raised the odds of all kinds of catastrophic scenarios—including the ones these companies are warning about.
I spent months investigating the frontier labs and AI "safety" experts seeking to regulate AI. What I found was a profoundly anti-human ideology that would alarm the average citizen and could determine the future of a world-altering technology.🧵
freebeacon.com/america/suici…
1
1
8
830
the ip was 64 hours 😭
PHASEONE[big]'s real name revealed to be PHASEONE64H?! Remember, [big] was a redaction by METR.
swarmtraces.org/viewer/#/row…
20
1,454
