Китайский новый год начнётся 10-го февраля

[dmm]

И это будет год "только что родившегося, растущего дракона"; а уходящий год - год "умирающего, уходящего кролика".

И тут-то, я чувствую, "всё" и начнётся; всё указывает на предстоящий "год критических потрясений", год, когда мир изменится радикально...

Хорошо бы нам его успешно пережить и войти в новую фазу...

Comments

[dmm]

https://en.wikipedia.org/wiki/Chinese_New_Year

https://en.wikipedia.org/wiki/Dragon_(zodiac) (note: dreamwidth has a habit of omitting ")" from urls, so click through at the Wikipedia suggestion)

https://en.wikipedia.org/wiki/Wood_(wuxing)

https://en.wikipedia.org/wiki/Rabbit_(zodiac)

https://en.wikipedia.org/wiki/Water_(wuxing)

2024 date : 10 February

2025 date : 29 January

[juan_gandhi]

Было бы неплохо, да.

[dmm]

10 months, 2 month till the next Chinese new year:

AI: everyone has GPT-4-like systems, they even exist in open weight configurations

o1 systems with slow thinking and reflection are the edge of a new revolution (Deep Learning revolution #4, 2012-2020-2023-2024)

multi-agent systems are formidable (e.g. we have 53% success on verified tab of SWEbench, OpenHands + CodeAct v2.1 (claude-3-5-sonnet-20241022), https://www.swebench.com/)

arcprize scores are formidable at 55.5 and 53.5 for two leaders

the ability of AI systems to do AI research is getting there (not "on par" with really strong humans yet, but getting close)

a lot of new super-promising start-ups: Ilya's Safe Superintelligence, Liquid AI, and so on

Politics: Trump won, and nominated a radical cabinet with the intention to dismantle the existing order

Wars everywhere: miserable nightmare, further escalations (Ukraine is pretty bad, Syria is starting to burn)

[dmm]

12 days of OpenAI (full o1, o1 pro mode, Pro plan, SORA, and other small change, 5 days to go)

A very impressive progress by Gemini.

Everyone is moving towards "agentic" (tool use), ready or not. Meanwhile, the ability of even relatively simple models to exfiltrate and copy themselves is demonstrated. Of course, various instances would be also able to exchange improvements.

Impressive results from many other orgs (Facebook, Amazon, ...).

Very impressive Chinese models (both open-source and closed-source).

Arcprize: a lot of materials published; plenty of ways to improve.

"Raw systems" are supposedly getting good on SWE-bench (but it's not clear where the ## come from). Meanwhile, the official verified leaderboard keeps inching up, now at https://www.swebench.com/: Amazon Q Developer Agent (v20241202-dev) 55.0, devlo 54.20, so added 2 points in 6 weeks.

GAIA benchmark got two new leaders in December, Langfun Agent v2.0 and Barcelona v0.1. Langfun is particularly impressive, https://github.com/google/langfun over Claude 3.5 Sonnet. Scores: Average/Level 1/Level 2/Level 3: Test: 49.33/58.06/51.57/25 Validation: 54.55/60.38/59.3/26.92. They had a particularly strong jump at Level 2, which finally became an approachable target with new LLMs.

[dmm]

о3

[dmm]

"Narrow AI" is likely this year (AGI-level artificial software engineer and mathematician capable of doing AI research on par with best humans).

Sam Altman says: "We are now confident we know how to build [narrow] AGI as we have traditionally understood it."

NVidia promises a revolutionary new laptop for $3000 or so: https://nvidianews.nvidia.com/news/nvidia-puts-grace-blackwell-on-every-desk-and-at-every-ai-developers-fingertips

Project Digits, a $3,000 personal computer powered by the Nvidia GB10 Grace Blackwell Superchip, includes a Blackwell GPU with 1 Petaflop AI compute and a Grace CPU with 256GB of high-bandwidth unified memory, claimed to be capable of running 200B-Parameter Models.

[dmm]

The next Chinese New Year is 29 January 2025, and on Jan 30 Sam Altman is allegedly doing this:

https://www.axios.com/2025/01/19/ai-superagent-openai-meta

"We've learned that OpenAI CEO Sam Altman — who in September dubbed this "The Intelligence Age," and is in Washington this weekend for the inauguration — has scheduled a closed-door briefing for U.S. government officials in Washington on Jan. 30."

[dmm]

DeepSeek-V3/R1 and associated events

[dmm]

And they gave people a technology allowing to convert any LLM into a reasoning model.

And Hugging Face started Open-R1 to reproduce the intricacies of their super-efficient training publicly.

Anyway, the Year of Dragon is over. The new year is the Year of Snake, and we are going to have an inflection in our AI takeoff this year.

o1 is a reasoning GPT2, o3 is a reasoning GPT3, o4 will be drastic (or even a good fine-tune of o3).

"Narrow AGI" this year: https://dmm.dreamwidth.org/87956.html