Krunch — AI High Signal

Nick Frosst @nickfrosst · 18:19

Technology is only good if it empowers you. If it locks you down, saps your attention, or steals your ideas, then it should be discarded. LLMs could go either way. That's why we are dedicated to open, sovereign, and empowering AI.

rohan anil @_arohan_ · 18:19

Every communication is a trade Typically, when you used a search engine or chatbot, you got a lot of value, while the service provider got your clicks, intentions, and chat data. Models have gotten significantly stronger and can now learn much more efficiently from very little

Baseten @baseten · 18:18

Speaker diarization is one of the hardest problems in voice AI and the best-performing models can be too slow and expensive to run at scale. We combined quality-aware quantization with index-based clustering to get @pyannoteAI 's models running 9.6x faster at 3x the throughput.

Adam Azzam a relayé

Adam Azzam @AAAzzam · 18:16

me in grad school: i wish biology was more like mathematics

Susan Zhang @suchenzang · 18:15

zoomer influencer shill: "there's no way they intentionally used user data for those results!" zoomer friends: "wow such high integrity to say that!" former cofounder: "akshualllyyy you can still identify someone pretty easily even if you anonymize everything, but don't worry!

Harrison Chase @hwchase17 · 18:15

I like this framing a lot: agent improvement is harness improvement, not model improvement The interesting interventions are often at the tool boundary: - what context you pass - when tools are provided - how you recover from failure - what gets measured afterward That’s the

François Fleuret @francoisfleuret · 18:11

Austin Lyons a relayé

Austin Lyons @austinsemis · 18:11

“Where do you see that extinction number landing at as we exit the calendar year?”

AK a relayé

AK @_akhaliq · 18:10

Tencent releases AuK, a unified speech generation and editing model It handles zero-shot TTS, instruction-driven content/acoustic/paralinguistic editing, plus enhancement and separation through one natural-language interface. The distilled AuK-Flash version runs 4.5x faster.

AK a relayé

AK @_akhaliq · 18:09

We’ve raised $47M in Series A funding led by @a16z to reimagine CRM as a world model of a business.

Aravind Srinivas @AravSrinivas · 18:09

Web app development usage (particularly those that require pulling information from several different places) is expanding quickly on Perplexity Computer. We're now supporting mobile and desktop previews for the websites created in the thread.

rohan anil @_arohan_ · 18:07

I think a few infinity stones do remain

kipply @kipperrii · 18:07

these wwii posters hit really hard because the memphis colossus datacenter is partially tva

Aidan McLaughlin a relayé

Aidan McLaughlin @aidan_mclau · 18:07

For the first time I am asking myself if things are moving too fast. I'm honestly not sure, but I am sure that it would be good for us to have an answer to "what would a successful pace look like?". I am hoping in the coming weeks and months a clear proposal is painted.

Joshua Achiam a relayé

Joshua Achiam @jachiam0 · 18:05

rohan anil a relayé

rohan anil @_arohan_ · 18:03

Mathematician Alain Connes, on the motivation for creating the Millenium Prize challenges: 'In contrast to Hilbert’s speech — and reflecting concern in some quarters that the CMI might be attempting a similar task — those responsible for the prize emphasize that they have no

Arena.ai @arena · 18:03

We analyzed both @claudeai Fable 5 and Fable 5.1 on writing clarity ranked by humans, and are sharing results later today. Which model do you think writes more clearly?

OpenAI Developers a relayé

OpenAI Developers @OpenAIDevs · 18:01

some @OpenAIDevs mail showed up at my doorstep :) myself and @purduehackers will be joining the codex ambassador program this fall! i’ve been using codex a ton already and am looking forward to building more with the team! lots of cool things to come, stay tuned :) and lots of

Julian Togelius @togelius · 18:01

New paper! We are honestly surprised that nobody seems to have explored this promising idea before. Results are interesting!

Yann LeCun a relayé

Yann LeCun @ylecun · 18:00

AI doomers are delusionally dangerous. Their solutions to an imaginary problem are: 1) Rigid authoritarian control over our lives and AI Or 2) Crash the economy by bringing progress to a grinding halt Again, all to solve a completely imaginary problem.

Julian Togelius a relayé

Julian Togelius @togelius · 17:59

What we’re witnessing right now in AI for math is a dramatic improvement in objective-driven problem-solving, but it does not speak to non-objective discovery. I have always maintained that objectives are achievable when they are one stepping stone away, which means that they are

max drake @max__drake · 17:52

come to our office in shoreditch next week and listen to me yap about some weird animation software we built

Lisan al Gaib a relayé

Lisan al Gaib @scaling01 · 17:51

Today we're launching Project Tailwind, a major call for founders to launch ambitious AI safety initiatives: We're looking for exceptional people to build the research, technologies, and institutions that will help humanity navigate the risks of

coefficientgiving.org Project Tailwind

max drake @max__drake · 17:51

the rigging system powering the animation of the Guy character in tldraw flash was quite complex to build, but it's allowed us a buuuunch of flexibility to create custom characters people can use in the app. it has too many sharpe edges to ship to users at the moment, but I would

François Fleuret @francoisfleuret · 17:50

What are some truths that will hold for sure in the "AI everywhere" universe?

Denny Zhou a relayé

Denny Zhou @denny_zhou · 17:50

Terry Tao is probably the most measured, pro-AI mathematician on the planet - which makes this quote especially concerning to read. 😟 If sharing your hunches means getting scooped ~immediately without acknowledgement, then we're going to see not just math but all other science

Visual Studio Code @code · 17:50

⌛ MCP Live starts in 10 minutes! Join @liamchampton @pamelafox and experts from Microsoft, GitHub, Anthropic, Okta, Prefect, and Amazon to explore what’s new across the MCP ecosystem 🎬

youtube.com 🔴 MCP Live!

Jared Zoneraich @imjaredz · 17:49

back in nyc

Ben (no treats) @andersonbcdefg · 17:49

really happy to see these guys get paid they deserve it

Kling AI @Kling_ai · 17:49

Some goodbyes don't come with a warning – but you can recreate them.

rohan anil @_arohan_ · 17:49

I think rivalry between ai lab leaders is fate being ironic otherwise collusion for maximum profit would put everyone else in bad spot

Tim Dettmers a relayé

Tim Dettmers @Tim_Dettmers · 17:47

We started to post videos for CMU 11-768, AI Agents! All of the videos will be posted to this playlist, so please bookmark/follow it if you want to be notified of new ones! I'll also try to post them on this thread too.

youtube.com https://youtube.com/playlist?list=PLSN0qpDfUvTM

EigenGender 🔸 ✈️ sf Sep 7-19 @EigenGender · 17:45

People on twitter will really be like 'you believe in working on alignment? that pales in effectiveness to my strategy, unplugging the datacenters' and then not unplug the datacenters.

Charlie O'Neill a relayé

Charlie O'Neill @oneill_c · 17:44

Some harness + training cooptimization we did with the cracked team @baseten. Tons of banger charts but my favorite is data room read coverage which moves from < 1% (base agent) -> 64% (RLM harness) -> 92% (RLM harness + RL) For basic behavioral strategies like "read the whole

Harrison Chase a relayé

Harrison Chase @hwchase17 · 17:43

You can use the LangSmith CLI to pull your flagged traces, classify each one by issue type, and build an eval dataset with a split for each. A 3 minute tutorial ⤵️

EigenGender 🔸 ✈️ sf Sep 7-19 @EigenGender · 17:41

it’s pretty scary that normal tweets on ai safety can blow up now. luckily I don’t have to worry about this due to a little trick called “my tweets are antimemetic” [bad]

EigenGender 🔸 ✈️ sf Sep 7-19 @EigenGender · 17:39

“One day everyone will have always been against this” but for accelerationism

Jürgen Schmidhuber a relayé

Jürgen Schmidhuber @SchmidhuberAI · 17:32

@BlackHackOfDoom Alas, McCarthy's AI had actually less to do with modern neural net-based AI than cybernetics and the much older adaptive neural net by Gauss & Legendre

Graham Neubig a relayé

Graham Neubig @gneubig · 17:32

We're releasing the videos for our AI agents course! So far, we've covered tool use, context management, skills and memory, and planning. Coming up in the next several weeks are domains (coding, GUI, deep research), and training methods (SFT, RL). Later on: frameworks,

Richard Socher a relayé

Richard Socher @RichardSocher · 17:32

The debate around recursive self-improvement is one of the most important, and contentious, in AI. I've heard people argue that the idea is nonsense: it’s mostly hype for impending IPOs. Lots of other people argue that yes, it's coming, but it’s dangerous, and we need to stop it.

Nicholas Dunzelman @nicdunz · 17:31

oh this is cool

merve a relayé

merve @mervenoyann · 17:29

we RL trained a 4B VLM to play GeoGuesser > environment, dataset, training recipe, evals, and code are all open-source. > runs on a single A100 > beat gpt 5.4 mini, haiku, qwen 3.5 122B and came close to sonnet on evals

Jimmy Lin @lintool · 17:28

Achievement unlocked! 🔓 Does this come with stat bonuses or help with saving throws?

uwaterloo.ca Jimmy Lin appointed Fellow of the Royal Society of Canada | Cheriton School of Computer Science | University of Waterloo

M Saiful Bari (MARUF) a relayé

M Saiful Bari (MARUF) @sbmaruf · 17:27

Anthropic employees

Alexandr Wang @alexandr_wang · 17:24

muse, built for moms ❤️

Jerry Tworek @MillionInt · 17:24

It’s time to call yesterday the beginning of the endgame. Midgame lasted about three and a half years. Needed new skills and few succeeded, but those that did won big. Bases are built out and tech is already advanced but the biggest discoveries are at the end of tech tree.

🇺🇦 Dzmitry Bahdanau @DBahdanau · 17:21

AI alignment will never be solved. Please plan the future of our civilization accordingly.

Cohere @cohere · 17:21

We're building AI to help people lead better lives. Do more of what you love with Cohere.

Kyunghyun Cho @kchonyc · 17:15

PSA:

Nicholas Dunzelman @nicdunz · 17:14

hmm... this doesnt seem to work too well for me. maybe im doing it wrong

Nicholas Dunzelman @nicdunz · 17:12

👀

Stanford NLP Group a relayé

Stanford NLP Group @stanfordnlp · 17:11

I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives. More thoughts below.

EigenGender 🔸 ✈️ sf Sep 7-19 a relayé

EigenGender 🔸 ✈️ sf Sep 7-19 @EigenGender · 17:10

@RBMD1982 I feel like people don’t think critically about how a couple dozen weird nerds posting on Less Wrong in the 2000’s accumulated power and influence at a rate ~ on par with the rise of Islam under Muhammad. Analytical thinking is very powerful!

Mat Velloso @matvelloso · 17:06

Glad to see an increasing number of people speaking up

Nicholas Dunzelman @nicdunz · 17:06

can you give it a finer brush and acrylics and bigger canvas and tell it to do smaller strokes and make a pointillism mona lisa

Nicholas Dunzelman @nicdunz · 17:03

all the little details are so good

Andrew Carr 🤸 @andrew_n_carr · 16:55

when everyone else is going big, go small?

Erin Price-Wright a relayé

Erin Price-Wright @espricewright · 16:54

Today, Covenant is emerging from stealth and introducing Anthem: a heavy-payload deep precision strike capability designed for mass production, with deliveries beginning in the thousands in 2027. We’ve raised $250 million, flown Anthem in more than 200 tests, and secured

Jamin Ball a relayé

Jamin Ball @jaminball · 16:54

Excited for Altimeter portfolio company @HQ_Covenant to emerge from stealth with $250m in funding and ~$150m in orders. Deep precision strike is one of the most critical capability gaps facing the West. Covenant has built Anthem to close that gap and meet the need for

Ryan Greenblatt a relayé

Ryan Greenblatt @RyanGreenblatt · 16:52

[Writing this in a personal capacity, not on behalf of my employer (Anthropic).] Jacob’s thread is very worth reading. Here’s my birds-eye view of the situation with risks from AI: 1. AI developers believe their technology could cause human extinction (or similarly bad

EigenGender 🔸 ✈️ sf Sep 7-19 @EigenGender · 16:52

can’t believe everyone is freaking out that Evan has a lower p(doom) than Lina khan

Google Gemini @GeminiApp · 16:50

Get your clutter under control with the help of Gemini Live. From shopping for new organizers to finding local battery disposal spots, Gemini can help you declutter smarter. Just open Gemini Live, point your camera at the mess, and talk through how to declutter it in real time.

sunil pai a relayé

sunil pai @threepointone · 16:48

"tbh I don't think developers want to experience anything at all" - @threepointone

sankalp @dejavucoder · 16:47

this website is one large message board and we are the agent swarms