๐๐ฐ๐ญ๐ฌ๐ด, ๐ ๐ฉ๐ข๐ท๐ฆ ๐ฅ๐ฐ๐ฏ๐ฆ ๐ฆ๐ท๐ฆ๐ณ๐บ๐ต๐ฉ๐ช๐ฏ๐จ ๐ ๐ค๐ฐ๐ถ๐ญ๐ฅ ๐ต๐ฐ ๐ต๐ฆ๐ข๐ค๐ฉ ๐บ๐ฐ๐ถ ๐ฉ๐ฐ๐ธ ๐ต๐ฐ ๐ต๐ฉ๐ช๐ฏ๐ฌ ๐ค๐ณ๐ช๐ต๐ช๐ค๐ข๐ญ๐ญ๐บ ๐ข๐ฃ๐ฐ๐ถ๐ต ๐๐, ๐ข๐ฏ๐ฅ ๐ข ๐จ๐ณ๐ฆ๐ข๐ต ๐ฅ๐ฆ๐ข๐ญ ๐ฐ๐ง ๐ธ๐ฉ๐ข๐ต ๐ ๐ฉ๐ข๐ท๐ฆ ๐ข๐ฏ๐ต๐ช๐ค๐ช๐ฑ๐ข๐ต๐ฆ๐ฅ ๐ฉ๐ข๐ด ๐ค๐ฐ๐ฎ๐ฆ ๐ต๐ฐ ๐ฑ๐ข๐ด๐ด. ๐๐ถ๐ณ๐ฆ
RESEARCH
-
Clarifying reported model ensemble scores on validation vs semi-private set
By
–
Sorry, but I think youโre misinterpreting the paper: The 62% is an ensemble of TTT + BARC on the public validation set and doesnโt imply any score on the harder semi-private v1. It shouldnโt be on that plot.
-
Neurosymbolic AI Development Overcomes Research Community Hostility
By
–
Completely messed up. I blame Hinton and LeCun for the counterproductive environment, likely delaying many key discoveries that could save lives by many years. Thank heavens @demishassabis didnโt let their unwarranted hostility towards neurosymbolic AI from letting him develop https://
t.co/bDmRMjnH3q -
Bridging Black Box AI and Explicit Knowledge Systems
By
–
as i know you would agree, we need a much better bridge between black boxes and explicit knowledge. we will see how long it takes. but i would agree that it is probably possible with current hardware.
-
Chain of Thought and RL in Extended AI Reasoning Analysis
By
–
the idea of analyzing things over a long derivation is about 60 years old, and will be part of the final answer; the particular way that o* does it (presumably via Chain of thought and RL but maybe other mechanisms) has not been transparently explained, and may or may not play a
-
AI Landscape Discussion: Reasoning, Hardware, Startups and Policy
By
–
daniel and i spent 2 hours together yesterday discussing ai over the last year – reasoning, bio, chips, startups, vibe shifts, policy, and ofc wassup @airstreet ๐
— Nathan Benaich (@nathanbenaich) 24 dรฉcembre 2024
episode dropping soon on @gradientpub ๐ https://t.co/jZhfxORjIHdaniel and i spent 2 hours together yesterday discussing ai over the last year – reasoning, bio, chips, startups, vibe shifts, policy, and ofc wassup @airstreet ๐ episode dropping soon on @gradientpub
-
Future AI Technology Evolution and System Integration Uncertainty
By
–
I wouldnโt at all assume that it will be. Maybe. Or if it is still there in recognizable form it might one module in a much more complex system.
-
FrontierMath Benchmark Transparency Concerns Raised by Community
By
–
Lots of questions around what was really shown in the FrontierMath benchmark, too. We need transparency, and instead got hype.
-
O3 unreliable across domains, better AI technology will emerge
By
–
Aside from details around ARC & other benchmarks, o3 wonโt turn out to be reliable across domains, and it will in hindsight seem wildly inefficient. MUCH better technology will emerge, after the LLM hype subsides. (LLMs may still exist, but will be just one tool in a larger
-
AlphaFold: A Greater Contribution Than ChatGPT
By
–
and, not unrelatedly AlphaFold is IMHO a much greater contribution to the world than ChatGPT