building ~ previously @SentientAGI @princeton_nlp

Joined November 2024
We’re excited to release @basiscompany’s first open source dataset: 1,500 hrs of track separated, multi-speaker conversations across 22 languages. The largest ever of its kind and first for several languages. Basis has built up a network of over 1 million contributors submitting audio, video, annotations, ratings, transcriptions and will keep open-sourcing interaction data to cover all the weird/ambiguous/natural modes of human interaction (more soon!). This release includes 70+ of hours with 3+ live speakers and 2645 native speakers meeting, sharing, arguing, laughing, trolling in English, Spanish, Japanese, Korean, Georgian, Mingrelian, German, French, Italian, Hebrew, Russian, Hindi, Arabic, Ukrainian, Dutch, Xhosa, Zulu, Portuguese, Chinese, Turkish, Polish, and Kazakh. For some of these languages, this is the first ever open release of duplex conversational data. 100 hrs are densely labeled with 113k human judgements on subtleties that models struggle with (was this backchannel sympathetic or frustrated? Was this silence awkward or turn-holding? Was this utterance directed or general? …) Reach out if you’re interested in more! ჯგირი მორაგადეს ჯგირი მარჩქილე ოკონია
13
9
3
35
9,831
We’re open-sourcing emotionally-steered speech across 108 countries/languages - by far the most of any steered audio ever released.  10,420 unique voice actors + 38 delivery styles (disgusted, flirtatious, whispering, singing, …)   525 hours of exceptionally curated, training-ready audio for expressive TTS. All audio reviewed by native speakers! We include all the usual higher-resource locales but also extend to a critically underserved long-tail of languages (Malagasy, Pashto, Oromo, Somali, Khmer, …) and dialects (Cypriot Greek, Kosovo Albanian, Libyan Arabic, …) where expressive speech training data is nonexistent. Reach out if you're interested in more hours / locales! (Link below)
4
6
1
35
1,185
Funny how speech-to-text models spam “don’t forget to like and subscribe” at the end of outputs… probably not supposed to train on that stuff! (Really common for Ukrainian, japanese, Turkish)
1
5
777
Like image gen producing watermarks
4
581
Fable just coined a brand new word and I’m genuinely so impressed. It invented “fushification” to describe a model overcorrecting valid masri (Egyptian Arabic) to fusha (standard dialect). I can’t find any mention of this word anywhere ever. Not even “fushify”. I can’t imagine the sub-word tokens are too friendly here either. Coining natural-sounding hapax is something humans do sometimes but it’s tough. At what rate do humans even come up with hapax (+ vs. models which apparently do too)? Real sign of intelligence…
6
3
13
689
Creston Brooks retweeted
AI indeed accelerates: the ones who think to think more and the ones who don't think to think even less
1
6
250
Creston Brooks retweeted
Are you worried that an LLM you trained could be stolen and misused by mysterious masked men 🥷? Our work (now a #NeurIPS2025 Spotlight 💫) can help you detect such unauthorized use. As a side-quest, we also analyse memorization and forgetting in LLMs 🧵(1/11).
10
3
2
31
8,126
Creston Brooks retweeted
Language models that think, chat better. We used longCoT (w/ reward model) for RLHF instead of math, and it just works. Llama-3.1-8B-Instruct + 14K ex beats GPT-4o (!) on chat & creative writing, & even Claude-3.7-Sonnet (thinking) on AlpacaEval2 and WildBench! Read on. 🧵 1/8
3
18
2
116
30,159
Creston Brooks retweeted
A16Z SPEEDRUN 2026 UPDATE: I think most people secretly know if they’re founders or not. Some of you can never be happy working inside a giant company, writing docs, in endless meetings. Deep down, you know you’re supposed to build. we're opening up a16z speedrun today! We are accepting applications for our 006 class, where we'll invest up to $1M. It's based in SF, kicking off Jan 2026 but you need to apply now in september. here's how to apply: speedrun006.a16z.com/x And yes, we are investing up to $100M in the next 30 days -- it's all happening in september. Quit whatever it is that you're doing, and in 2026 come to SF and work with us out of Andreessen Horowitz's office in SOMA, alongside over a hundred other founders, building the startup you've always wanted to build. We will help you -- both myself and the a16z speedrun team. the details: - up to $1M of investment - hosted at a16z HQ in San Francisco - 12 week program, with an IRL kickoff, luminary speakers, community events - live events with the founders of Carta, Zynga, DoorDash, Behance, Airtable, Twilio, Figma, and more - private dinners/Q&A with Marc and Ben of a16z - apply now, and the deadline will be Sep 28 2025 for SR6 At a16z speedrun, you get access to programs from our operating team and work with experts in marketing, BD, talent, people and capital—more below MARKETING Our team of expert marketers is here to help you win. Whether refining your brand, launching, or building a thriving community, our marketing operators have powered dozens of startups with: - Brand Development -End-To-End Marketing Strategy -PR & Media Coverage -Go-To-Market Execution -Creators & Content TALENT Find and attract the talent you need to build and scale your company. Our curated network connects you with world-class technical talent, executives, advisors, and specialists who can help accelerate your success. Here’s how the program works: - You tell us what you’re looking for. - We use a16z speedrun's brand and referral networks to magnetize talent. - We take hundreds of calls each week to curate a list that we only send to speedrun founders. - You request introductions and we put you in touch. PEOPLE We help you quickly stand up the tools and practices needed to hire, manage, and lead highly performant teams. Our goal is to help you anticipate challenges and navigate some of the most foundational decisions you'll make as you build a world-class company. While a16z speedrun takes place in the US, we welcome founders and companies from around the world. Our Global Founders Program provides specialized guidance for navigating visas and relocation, plus dedicated access to our expert immigration attorney network, so you can focus your energy on building your company. BUSINESS DEVELOPMENT We unlock unparalleled access to networks, expertise, and tools that help startups scale faster. - $5M+ in free credits in our speedrun Marketplace from AWS, GCP, OpenAI, Anthropic, Microsoft, NVIDIA, Stripe, Deel, and many more. - Dedicated advisors & experts to guide you through every partnership, sales, and GTM motion, including crafting your business model and pricing. - Level up your sales with real live demo experience across various events to executives and operators in your target market. - Access to a16z’s network of executives & decision-makers at 2,000+ companies. CAPITAL One of the most important things a founder can do is raise money. Our fundraising program sets the stage for your raise through an in-person Demo Day and an online platform reaching 1,000+ top early-stage investors. We help you prepare with practice sessions, stress tests, and materials review. When you’re in high-stakes negotiations, we coach you, share insider knowledge, and leverage alumni intel on the investor across the table. More details: The a16z speedrun program is a fast‐paced, 12-week startup program that guides founders through every critical stage of their growth. It kicks off with an orientation to introduce the cohort, then dives into rapid product development—helping founders think through MVP while addressing key topics like customer acquisition and design partnerships. Throughout the program, startups benefit from expert-led sessions and interactive office hours that cover: - Brand Building & Go-to-Market Strategy: Crafting your story, marketing, and driving product-led growth. - Customer Acquisition & Launch: Securing early users and executing effective launch plans. - Fundraising & Strategic Partnerships: Pitching, navigating investment, and building lasting relationships. - Team Building & Operational Scaling: Developing high-performing teams and refining internal processes for sustained growth. - Community & Enterprise Sales: Building communities, forming strategic partnerships, and landing your first enterprise customers. - Product-Market Fit & Demo Day Prep: Assessing market traction which culminates in a Demo Day to showcase progress. The a16z speedrun program is IRL and runs for 12 intensive weeks in which our team of expert investors and operators guide your startup from idea to market launch. The program moves through sequential modules—each dedicated to key aspects such as rapid product development, go-to-market strategy, fundraising, team building, and operational scaling. Expect regular check-ins, one-on-one office hours, and interactive sessions, culminating in a Demo Day where you present your progress to potential investors.
118
114
64
1,030
5,940,587
GPT-5 eval’ing on collie, tau^2, swe bench, charxiv… big from @princeton_nlp :)
10
751
Truth
Sometimes you have to just say fuck it and let shit go because stressin out over things you can't control is a waste of time 👌🏼👍🏻
10
1,189
moved to LA / building smth cool with @AlexisS1776 & @speedrun
2
13
1,244
The line between unhinged and cringe in b2c self promo is extremely thin
1
1
9
1,057
Feels like ~1/3 of reviews on icml / colm / neurips etc these days are pure gpt. There must be a decent number of papers who win the reviewer lottery getting published based only on LLM vibe checks :)
2
871
can @GeneralTxn translate UI into boomer ty for the help
1
5
813
sf formal from the @GeneralTxn box lmao
2
673
Crazy that (afaik) no LLM has solved a 3x3 Rubik’s cube even with fine-tuning (generalize much?). At the same time, action space is small and the reward is simple (solved or not)… seems ripe for LRMs / RL. Some thoughts as a long-retired cuber…
2
1
6
968
Another thing to try would be a sort of curriculum learning approach… first teach cross solving, then F2L, then LL. Obv this is less pure but potentially more feasible / might converge quicker (no need to invent a whole new method). The other cool thing here is the potential to find things new techniques that might be useful to modern CFOPers (or ZBers, etc.). Like, add some additional reward for minimizing rotations, maximizing R U moves (or whatever the TPS meta is).
1
1
462
Maybe even a move-count penalty would discourage alg-heavy methods — like, to discover new algs the model will have to mess up a bunch and then have start the solve over, incurring heavy penalty). Maybe algs would have to be some emergent thing where a model has to have gained insane spatial awareness that it can just first try find a new alg — or the reasoning trace has to be long enough to do a sort of local tree search through potential moves.
1
360