Aleks

@ZhengyangGeng Best photo of the event lol

When you see a semi-open laptop at an ICML event 😂

@MetatrolN @fin_presented Great paper! Keep it up! Btw @theo check this out lol

AI4Math research focuses on Lean autoformalization and theorem proving. But how do we come up with new conjectures? In our work with @fin_presented, we study how well can LLM agents use computation to create hypotheses and solve research-level mathematical problems? arXiv: arxiv.org/abs/2607.06820 We evaluate 15 LLMs in zero-shot and SageMath-augmented agentic setup on RealMath 133 problems extracted from arXiv mathematical papers. Key findings: - Tool access improves every model, by 9.7 pp on average. Open-weight models gain 15.3 pp, compared with 6.5 pp for closed frontier models, narrowing the gap between them. - Most intriguingly, a CAS-augmented agent reproduced a computational mathematician’s workflow: computing intermediate objects, finding patterns, forming conjectures, recovering from errors, and validating formulas across parameters (see the details in the case study). - Gains vary: #Qwen 3.7-Max rises from 42.1% to 69.9%, a gain of 27.8 pp that brings it close to frontier performance. #Kimi 2.7 gains only 1.5 pp. - Tool-use behavior is strongly bimodal. Strong agents usually finish in 3-4 tool turns, while weaker agents often exhaust all tool budgets. - The largest gains are in combinatorics (+18.7 pp) and rings and algebras (+10.7 pp), while algebraic topology and group theory remain difficult. - Recovery after a failed tool call ranges from 16% (#Sonnet-5) to 77% (GPT-5.5) across models. The ability to revise a strategy after receiving computational feedback separates effective agents more clearly than the raw number of errors. Interesting observations about some models: - #GPT 5.5 leads in both solve rate and efficiency, reaching a 75.2% accuracy with the lowest token usage among tool-enabled agents. - #MiniMax M3 is the least efficient, using the most tokens per problem and achieving substantially lower accuracy. - #Opus 4.8 exceeds Opus 4.7 by only one solved problem. - #Grok 4.3 shows one of the worst results and produces 248/336 SyntaxErrors🥲. - #Fugu Ultra shows the smallest increase in token usage with tool access, at 4.5×, averaging 70k tokens per problem. Today we are presenting our poster at the @ai4mathworkshop at @icmlconf. Come by to discuss our work. #ICML2026 #AI4Math #AI #Agentic #LLM #Mathematics

@IRCSS Insane amount of work, just wow

My child was born so I have been really busy. But I managed to clean up the magic school demo scene and add it to the repo. So far everything was generated using 100 percent geometry nodes. Now its time to add in some hand modeled stuff github.com/IRCSS/Blender-…

@vaxryy @temidaradev Goated album

So @temidaradev I tried your music player and it's alright, but I have some issues: - Tidal integration (their API is a bit ass but I'm sure you can manage) - playlists, searching, playback - UI optimization - this is the laggiest desktop app UI-wise I've ever used even including electron. Scrolling is legit 20fps. - Listenbrainz fixes - "listening now" is sent but the listen is rarely registered afterwards (like 25% of the time?) - likely HTTP errors, LB's API is notoriously finnicky - need to check for 429 and timeouts. Because I like seeing open source grow and because dara is a nice person unlike a certain other turkish kid, I'll put my money where my words are: I'll donate 100$ each for the first two, listenbrainz is probably easier so 50$. Posting publicly so y'all can rip my nuts off if I don't pay afterwards. (also as a PS. the settings menu is quite hard to navigate, more clear sections, less spacing between options, and clear distinction where one ends and another begins would really help)

@theo Doesn't really work well with topology and graph theory in my experience

I'm looking for people who have given Fable a real shot and are unimpressed. Tell me more about your experience with it below. Will follow up with a few specific questions

@KenneyNL Notepad++, duh

People were a little worried that I was using MS Paint for pixel art before switching. I'm almost afraid to tell I'm using Notepad++ for coding, made whole MakeRoom using it... Any better recommendations? I don't want any bells and whistles, just a super basic IDE. No AI, etc.

@nikitabier Sorry for your loss

My mom passed away last night with 10 family members by her side. She was the first in our family to escape the Soviet Union. She arrived in California with nothing, hustled as a single mom of 4, and achieved the American dream in every respect. She was crafty, resourceful, and could make humor out of any situation. Here’s a photo of us when money was tight: I asked to be Aladdin for Halloween, so she dressed me up as a Russian gypsy and thought I wouldn’t notice. April 20, 1950 - June 21, 2026

Community note of the year

@LPLive @ChesterBe Last year's Frankfurt night 2 clears, unfortunately

THE BEST LINKIN PARK OUTRO EVER...? Take it away, @ChesterBe! Faint, live at Rock am Ring 2007

@geogristle When's the preprint? Looks really cool!

It works

@DiscussingFilm Wait why does this look like a WWE guardians of the galaxy spinoff lol

New clip from the ‘MASTERS OF THE UNIVERSE’ movie. In theaters on June 5.

@AL33012 1) what

国際宇宙ステーション #ISS からの電波、三重県津市から液体アンテナでも取れました！ CMの都合で高度の低い18:30頃までの受信でしたが51でしっかり復調しました👍 コップの水位は14センチ位で2m のλ/4には満たないですが波長短縮効果である程度のマッチングにはなっていたと思われます♪

@HidekiKamiya_X is my favorite twitter account by far these days. Thanks for the laughs man, I wish I could come up with such exquisite curses

Did they just use THIS song for the android announcement?

@scheminglunatic don't forget nvidia drivers

@konekone2026 @HugoCGuerra Could you please add support for blocking custom websites? (And maybe custom cats?:))

Cat Gatekeeperの公式サイトを作りました！日本語と英語対応です。今後の新しいアップデート情報など、こちらからご確認ください💁‍♀️

@scheminglunatic my favourite shitpost

@theo If u tell me what you need to be added I'll just send you app bundle (or code, whatever)

@theo Hey, Theo! It seems current version of OBS "introduces" pretty big latency and "soapy" picture. I know this cause I tried using it with my low-end capture card and got results so bad I went on to write custom soft for it.

Bounty: I will pay you $10,000 if you find me a capture card that does all of the following (READ TO THE END BEFORE YOU REPLY OR I AM BLOCKING YOU) Required specs: 1. Can receive 1440p 60fps signal over HDMI from a Mac (in 720p hidpi mode) 2. Capture works on Windows over PCI-E or USB 3.2 3. Capture works in OBS 4. Video format contains options that prevent grays from flickering (XRGB works usually) 5. Doesn't randomly disconnect if PC is left on 6. Audio passthrough works reliably Cards I have tried that do not work: Avermedia GC553G2 This one is really really close. Only one I've found that works with hidpi and has no color flicker issues (on XRGB). Sadly it disconnects ALL THE TIME when I leave my PC on, and sometimes fails to reconnect entirely, forcing me to reboot. 50% of the time it "looks fine" but fails audio passthrough. SO CLOSE, but unstable enough to make me crash out. Elgato 4k X The spec sheet says it will work. @Videophile says it will work. It will not work. Sometimes it does for a single stream, but then it refuses to honor the color encoding settings and the flicker returns. There's also a bunch of settings they just won't let you change in OBS? Idk, I tried for days to make this card work and gave up. (do not recommend other elgato cards, this is the only one that meets my spec requirements and it still doesn't work) EVGA XR1 Didn't have resolution or color options I needed. Impressive for the price but not viable for my work. BlackMagic Decklink Quad HDMI 4k This card has taken years off my life. It only accepts "cinema resolutions" from my Mac. It "supports" 1440p, but will not ingest it no matter what I do. Unless I send it from a PC instead of a Mac. Then it works fine. For no reason. Ugh. Also missing the color options I need sadly. Magewell USB Capture HDMI 4K Pro 32150 I don't know if this will work or not. I already ordered it and it's on the way. I am trying not to get my hopes up. How to qualify for the prize 1. Recommend me a capture card that has not been recommended already in the replies. 2. Show the specs I need documented publicly somewhere 3. Wait patiently for me to order the card, test it, and report back You will not be paid until I use the card for 2 weeks without issue. If there's some speed bumps initially that's fine, I just need 2 weeks of pushing it without issue to know that "i'm finally good". This has been bottlenecking me for 5 years now. I am so tired. Please help.

@Polymarket Reach out and touch faith (C)

JUST IN: New app lets users pay $1.99 per minute to talk to an AI-generated Jesus.