AdamK

204 posts

AdamK

AdamK

@AdamK133

ML Safety. Interested in AI for epistemics. Formerly: Cal.

Katılım Ocak 2022
289 Takip Edilen95 Takipçiler
AdamK
AdamK@AdamK133·
"Experiments"? Oh, you mean cost-sensitive tool calls.
English
0
0
0
14
AdamK
AdamK@AdamK133·
@Sauers_ Mech interp may have huge capabilities externalities.
English
1
0
3
348
Sauers
Sauers@Sauers_·
Any computation graph can be realized as a neural network. Good compilers optimize computation graphs to eliminate redundancy. We currently have no such compilers for neural networks. Why?
English
11
1
48
4.5K
AdamK
AdamK@AdamK133·
453 days. (To be fair to @EpochAIResearch I think the actual FM T4 turned out well.) I do genuinely wonder what there will be left to benchmark in AI math 453 days from now, besides human math's very toughest open problems.
AdamK tweet media
Davis Brown@davisbrownr

Mathematicians use conjectures to point to important, open problems. We collect nearly a thousand (currently 890) recent conjectures from the math literature for a new dataset, OpenConjecture. On a subset, GPT-5.4 finds candidate proofs, and formalizes several in Lean.

English
0
0
2
88
AdamK
AdamK@AdamK133·
The core abstractions of the coming paradigm are taking shape.
Paradigma@paradigmainc

Flywheel views Directed Acyclic Graphs as the underlying data structure for research: nodes can be observations or experiments. When they're experiments, they contain an hypothesis and the artifact of the experiments that may reject it. In Flywheel, researchers can focus on their ideas and allow agents to implement code, provision compute, and structure their experiments and results, while collaborating with other agents and scientists.

English
0
0
1
42
AdamK
AdamK@AdamK133·
I too am concerned about short-term AI water use!
English
0
0
1
93
AdamK
AdamK@AdamK133·
@forethought_org has had the highest-quality past year of conceptual research output of literally any org in recent memory. It's not even close; bravo.
English
0
0
4
73
AdamK
AdamK@AdamK133·
The Great Legibilization is well underway
English
1
0
0
69
AdamK
AdamK@AdamK133·
They say you have the average AI timelines of the five people you spend the most time with.
English
1
1
7
200
AdamK
AdamK@AdamK133·
This is a good thread that would approach the heart of the matter, if only it changed one word. Mathematics is a fundamentally *instrumental* subject. But a mathematical idea's instrumentality, both inside and outside the subject, does not need to be defined anthropocentrically.
Jonathan Gorard@getjonwithit

@stephen_wolfram What it made me realize is that mathematics is a fundamentally *human* story. The Babylonians developed arithmetic for commerce. They developed geometry to survey land. We generalized arithmetic to get equations, then abstract algebra. We generalized geometry to get... (10/15)

English
0
0
0
66
AdamK
AdamK@AdamK133·
In retrospect, no surprise that Elon is using transformers to turn his car company into a robot company…
English
0
0
2
85
AdamK
AdamK@AdamK133·
@xeophon Fwiw I tend to be one of those people, but I only think it’s required if models are still asymptotically quadratic in context length. I also acknowledge that standard long context windows + better (possibly model-directed) context management will give v large effective contexts.
English
0
0
0
76
Xeophon
Xeophon@xeophon·
I bet the people who say that c*ntinual learning has to be done in weights also would’ve said that reasoning has to be done in latent space before o1/r1
English
25
11
281
34.1K
AdamK
AdamK@AdamK133·
@dpaleka 4. Some process to move skills and knowledge into the weights once their importance is well-validated across many tasks.
English
0
0
2
96
Daniel Paleka
Daniel Paleka@dpaleka·
here is how to solve continual learning: 1. long context 2. all useful user or env feedback goes to claude md 3. when a section in claude md is long, it becomes a skill that's it, two levels of hierarchy, no finetuning, don't overcomplicate things. x.com/bcherny/status…
Boris Cherny@bcherny

4/ Our team shares a single CLAUDE.md for the Claude Code repo. We check it into git, and the whole team contributes multiple times a week. Anytime we see Claude do something incorrectly we add it to the CLAUDE.md, so Claude knows not to do it next time. Other teams maintain their own CLAUDE.md's. It is each team's job to keep theirs up to date.

English
9
15
229
79.2K
Ryan Greenblatt
Ryan Greenblatt@RyanPGreenblatt·
It seems pretty likely that AI has large effects on the economy and massive effects on software engineering well before coding at AI companies is fully automated. So 2027 might be pretty crazy while full automation of AI R&D is still at least several more years away.
English
7
2
78
7.5K
AdamK
AdamK@AdamK133·
DAOs of Loving Grace
English
0
0
0
66
AdamK
AdamK@AdamK133·
When the aesthetics of forecasting win the day, "life expectancy" will finally refer to the average age a person born today can expect to live. Hell, there'll be a distribution too. I wonder what people will think of the early hump and the long tail...
English
0
0
2
90
Daniel Eth (yes, Eth is my actual last name)
Daniel Eth (yes, Eth is my actual last name)@daniel_271828·
Bernie becomes the first member of Congress to support an AI pause. Prediction - more to come as the issue increases in salience
Sen. Bernie Sanders@SenSanders

I will be pushing for a moratorium on the construction of data centers that are powering the unregulated sprint to develop & deploy AI. The moratorium will give democracy a chance to catch up, and ensure that the benefits of technology work for all of us, not just the 1%.

English
9
2
54
2.7K

Keşfet

@Sauers_ @EpochAIResearch @peterwildeford @forethought_org @xeophon @dpaleka @elonmusk @BarackObama