Noise

Hoping AGI replaces my job first.

[Noise @llmwhispers]

You need to be job hopping. You should have your LinkedIn open waiting for a recruiter to slide into your DMs offering a 50% salary raise. No company is going to offer you that kind of raise after a year but the market will. Forget people like this, they just want cheap labor.

[Brett Adcock @adcock_brett]

I often see resumes where someone has worked at 6 companies over the last 10 years.

I'm staring at one right now. These job-hoppers usually last a little under 2yrs per company on average.

If someone has a jumpy resume, it's the first thing I notice. It's so blatant to me, yet it continues to fool a lot of people.

Thinking back over the last 15yrs at my companies:

I've hired thousands of people, and I can't think of a single person I've hired with a jumpy resume who has worked out long-term. Not one.

[Matteo Capucci @mattecapu]

never laughing at crackpots again

[wordgrammer @wordgrammer]

Chat GPT is really good at solving math puzzles and very bad at handling Cmake. Unfortunately, I want to do math puzzles, and leave Chat GPT to handle Cmake.

[Alex Kontorovich @AlexKontorovich]

Oof. Remedial algebra. At Harvard. I wish I could say I'm surprised... Very sad state of what constitutes education... We have to keep pushing to do better.

[imit @imitationlearn]

wait so apparently 4chan figured out step-by-step reasoning as a emergent property of gpt-3?!

[Noise @llmwhispers]

@Chenlu0101 @stringking42069 About a decade ago the Simons Foundation threw huge sums of money to theorists and ruined everything.

[Noise @llmwhispers]

@ThomasVanRiet2 What is the argument for N=3 being large?

[Alex Kontorovich @AlexKontorovich]

In just a few years, this will no longer exist. Math books will be formalized, and any "obvious" statement will contain a hyperlink to a proof that you can interrogate down to the axioms. It will be glorious!

[anton @abacaj]

google finds a way to make me bullish on small models again

"... this indicates the reasoning performance of the small LMs may be improving more rapidly over time compared to the larger LMs"

[Jim Fan @DrJimFan]

Introducing NVIDIA Cosmos, an open-source, open-weight Video World Model. It's trained on 20M hours of videos and weighs from 4B to 14B. Cosmos offers two flavors: diffusion (continuous tokens) and autoregressive (discrete tokens); and two generation modes: text->video and text+video->video.

Physical AI has a big data problem. Synthetic data to the rescue! We apply Cosmos to large-scale synthetic data generation for robotics and autonomous driving, and now you can too! It's all yours to finetune.

Check it out: https://t.co/ZGCeWRd2vj

[Noise @llmwhispers]

@Yi_Zen_Chu @NipseyHoussle Because the supply of available people willing to work at 50k would change, either the demand for the work would change or the compensations/postdoc lifestyle would have to in order to attract the domestic talent that go to greener pastures and not do postdocs

[Noise @llmwhispers]

@Yi_Zen_Chu @NipseyHoussle The minimum 300K would fix the same problems in academia that it would fix in industry: employers hiring cheap labor in mass to undercut US wages.

[anton @abacaj]

Incredible, still going

[Noise @llmwhispers]

A gentleman and a scholar with his posts.

[Bahram Shakerin @BahramShakerin]

I would like to send my best wishes to everyone celebrating Christmas and Hanukkah. Wishing you and your family all the best.

[Bahram Shakerin @BahramShakerin]

Field Redefinitions Can Be Nonlocal

Timothy Cohen, Matthew Forslund, Andreas Helset

https://t.co/5IFqsHk2Zl

[Daniel Green @nu_phases]

Dec 12 - A de Sitter S-matrix

https://t.co/85IXPNnZEJ

Marolf et al. define an S-matrix for massive particles in (fixed) global dS. They prove many facts about flat space S-matrix apply to dS. They develop a lot of interesting ideas that should have a wide range of applications

[Daniel Whiteson @DanielWhiteson]

Periodic reminder

[Marcos Mariño @marino_beiras]

Today I posted to the arXiv the written version of my lectures at Les Houches on non-perturbative topological strings. Please send me your comments and/or complaints 😊

https://t.co/s0SZimFoAi

[Greg Brockman @gdb]

Interesting small-scale study on accuracy of diagnosing illness:

- Human doctors: 74%
- Human doctors using ChatGPT: 76%
- ChatGPT alone: 90%

Takeaway seems like vast potential for AI to help with diagnosis, but need better human <> AI teamwork:

https://t.co/evE1jurR3U

[Surya Ganguli @SuryaGanguli]

Some numbers about language experience:
LLMs are trained on > 1T words.
We get 100M words.
It would take us 24,000 years to read 1T words.
Our DNA is ~700MB or 6B English words.
Conclusion: info from *both* learning and 4B years of evolution is *far* less than what LLMs get.