Polo M (slow/steady) @traintest_split

Research Scientist. AI, machine learning and tech policy. after hours ai takes. Joined January 2023

Tweets

5K
Followers

100
Following

950
Likes

3K

机器之心 JIQIZHIXIN @jiqizhixin

15 hours ago

OpenAI just claimed to have found the culprit behind AI hallucinations.

105 346 4K 373K 2K

Download Image

LLMs act sub-optimally in decisions due to greediness, frequency bias, and a knowing-doing gap. A classic @GoogleDeepMind paper. Shows why LLM agents make poor decisions and how reinforcement learning fine-tuning fixes a chunk of it. Tic-tac-toe win rate jumps 15% to 75% after…

4 45 252 14K 211

Download Image

andrew gao @itsandrewgao

19 hours ago

i had to prompt inject the @united airlines bot because it kept refusing to connect me with a human 🧵 what led up to this breaking point

231 890 24K 2.2M 8K

Download Image

Polo M (slow/steady) @traintest_split

a day ago

long term memory, abstraction, and context remains critical

Warren Hatch @wfrhatch

a day ago

long term memory, abstraction, and context remains critical

3 4 11 721 2

0 0 0 13 0

Victor @victor_explore

2 days ago

Algorithmic breakthroughs are the most important factor. With algorithmic advances, you don't need as many parameters or training steps

1 3 3 443 2

Download Image

Oren Cass @oren_cass

2 days ago

2/ The Banks bill takes the sensible and modest approach of requiring US chipmakers to offer AI chips to American customers before selling them to China. So it's just blocking sales where more chips for the CCP directly means fewer for American firms. theregister.com/2025/09/04/us_…

2 7 61 6K 4

Polo M (slow/steady) @traintest_split

2 days ago

futurestructure coded

Nabeel S. Qureshi @nabeelqu

2 days ago

futurestructure coded

0 0 4 629 0

0 0 0 7 0

Matt Shumer @mattshumer_

2 days ago

This is the correct take. Evals are helpful but not well-correlated with actual utility. At Otherside, we use A/B tests grounded in real-world traffic, measured against subscriptions and retention. We've tried it all. This is the way.

ben (is hiring engineers) @benhylak

2 days ago

9 3 90 19K 37

1 3 28 8K 10

Ethan Mollick @emollick

2 days ago

This paper finds LLMs' ability to understand that others have different beliefs (Theory of Mind) comes from 0.001% of their parameters. Break those specific weights & the model loses both its ability to track what others know AND language comprehension. Interesting implications.

39 134 1K 75K 694

Download Image

First Squawk @FirstSquawk

2 days ago

ANTHROPIC TO HALT AI SERVICE SALES TO MAJORITY CHINESE-OWNED COMPANIES: FT

3 4 22 9K 5

Tom Dörr @tom_doerr

2 days ago

mesh VPN built on WireGuard for easy secure device connections

2 2 25 3K 21

Download Image

Tenobrus @tenobrus

2 days ago

all i want is 15-30 highly intelligent autonomous drones with fine motor manipulation capabilities i can direct to pursue my goals and take action in the world on my behalf night and day . is that too much to ask ????

20 4 168 4K 12

Kol Tregaskes @koltregaskes

2 days ago

Warner Bros. Discovery sues Midjourney, alleging the AI tool generates countless unauthorised copies of its characters. - The complaint says Midjourney “brazenly dispenses” copyrighted IP as if its own - Alleged unauthorised derivatives via image and video generation tools -…

The Verge @verge

2 days ago

2 8 45 22K 4

4 2 10 4K 3

Download Image

Enrique Ide @ide_enrique

3 days ago

So what do we find? Key lesson: AI breaks the neat “substitute vs. complement” story. Depending on how it’s used, AI can substitute some and complement others. (6/n)

1 8 22 2K 4

Download Image

Damien Masson @damienhci

3 days ago

Visual Story-Writing. While you write, our word processor visualizes the timeline, world map, and character relationships. Editing these visuals updates the story (e.g. drag a character on the map to move them). This summarizes our #UIST2025 paper. #HCI #LLMs #AI Thread 🧵 (1/8)

104 291 2K 330K 3K

Download Video

Cole McFaul @colemcfaul

3 days ago

🚨 NEW REPORT! 🚨 After 2 (!!) years working on a new dataset of thousands of PLA contracts, I’m very excited to share a new report, featured in today’s @WSJ! We ask: Who are the PLA’s AI suppliers? The answer: an unexpected, emerging class of firms and universities. 🧵/15

19 158 500 71K 316

Download Image

signüll @signulll

4 days ago

in a wild plot twist, openai’s rise literally saved google from being broken up. absolutely unreal. you couldn’t script this if you tried.

82 278 3K 302K 1K

Download Image

Peter Wildeford🇺🇸🚀 @peterwildeford

5 days ago

🧵 1/13 Compute is the new oil. It's cliché but true, especially geopolitically. Like oil, compute is: - Scarce (demand far exceeds supply) - Concentrated (US: 75%, China: 15%, EU: 5%) - Controllable (export controls actually work) - Strategically vital (biggest AI bottleneck)