Deedy

@deedydas • 243,238 subscribers

Partner @MenloVentures. Formerly founding team @glean, @google Search. @cornell CS. Investor @AnthropicAI, @GoodfireAI, @OpenRouter, @WisprFlow, @_inception_ai

Live Cam

Deedy

Shorts

Billionaire Michael Milken joked “if a US company replaces the US-born CEO with a CEO born in India, I buy the stock” But he reveals he hasn’t backtested the idea. So we did. In the last 15yrs, that would’ve 50x’d your money: 7.5x more $$ and &gt;2x IRR vs S&amp;P500: 30% vs 14%!

Wow, Gemini 3.1 Pro just crushed one of my test problems no other model could solve. "Write a photorealistic 3D ocean simulation" 3D graphics is extremely hard. It has to encode 9-10 techniques based in physics. Being off by just 1 number can lead to bonkers results

You can buy the humanoid Unitree G1 for $16,000. 4'4" tall, 35kg, 2hr battery life. This is not Boston Dynamics. This is a real thing you can buy.

As cool as the new Sora is, gymnastics is still very much the Turing test for AI video. 1/4

We JUST solved a hard open math problem— the moving sofa problem! "What is the area of the largest shape that can be manuevered through an L corner of width 1?" On Dec 2, 2024, Korean Jineon Baek proved the answer is the 18 curve section Gerver's sofa with area 2.2195..!

Grok just dropped Companions, animated AI characters that talk to you in realtime! There are 2 so far: — Ani, a sexual anime girl who wants to cuddle and — Bad Rudy, a mischievous fox that just wants to cuss you out. Judge for yourself, here’s a video:

Two 23 year old Indians just dropped the #2 open-weight AI voice model in the world, trained purely on free credits! Maya1 is #20 globally, better than even Google's best. 3B params, runs on one GPU and does 20+ emotions with &lt; 100ms latency You can just do things.

This new Gmail smart scheduling feature is a straight +4% to global GDP

Llama 4 seems to actually be a poor model for coding. Scout (109B) and Maverick (402B) underperform 4o, Gemini Flash, Grok 3, DeepSeek V3 and Sonnet 3.5/7 on the Kscores benchmark which tests on coding tasks like this. ELO-maxxing on LMarena doesn’t create the best models.

Bytedance just dropped realtime voice translation 3x faster than before, with only a ~3s lag! Seed LiveInterp 2 is a full duplex speech-to-speech model with &gt;70% correctness. When this makes it to video calls, it'll open up previously impossible connections.

It's so over. AI can now control your iPhone and chat with all your dating app matches!

Here’s Bad Rudy:

Sonnet 3.7 generated this entire "AI agent startup" website from scratch. It's superpower is its 64k *output* token length, which makes the resulting Artifacts way richer. This is better than most actual startup websites.

Cute little sign in page concept which watches you type and closes its eyes when you're typing the password. Entirely on Claude Sonnet 3.7.

I made Computer Use work for MacOS directly and let it loose on my actual computer!! Anthropic's "safe" Ubuntu docker demo was too painful. In this video (8x), it creates a Double Pendulum simulator from scratch and opens it up in Chrome. Put it all on Github too. Feel the AGI.

There's a new foundational AI model that specifically uses computers fast. It's called Ace. Most similar things retrofit heavy vision and reasoning models which are slower than a human. Watch it book an Airbnb in 15s!

Videos

Watch Anya Live

Gymnastics is the Turing test of video generation models

Claude Code is one of the worst named products. It's not just for code. Watch this Hermès ad. 30 seconds. 8 shots. Voice. Music. Even text branding burned in with ffmpeg. Clopus 4.5 wrote a script, orchestrated ElevenLabs, Google Veo 3, downloaded music and made this from scratch.

defense tech is so freaking cool

Neuralink's latest demo actually brought me to tears. You can use your brain to play Mario Kart, Call of Duty and even control a robotic arm to write. As someone with an uncle who is disabled, it's awesome to see a team truly unlocking human potential.

Karpathy's Autoresearch pushed my vibecoded Rust chess engine AI from "expert" to a top 50 grandmaster, a #311 chess engine. It ran over 70 experiments on its own and tried to hill climb to the top ELO score it could, landing at 2718!

Blows my mind that we currently possess the technology for Google Maps to turn all the street view images of the entire world into a video game you can play! In the future, we'll be able to say "yeah let's check out New York City 100 years ago!"

Manus, the new AI product that everyone's talking about, is worth the hype. This is the AI agent we were promised. Deep Research+Operator+Computer Use+Lovable+memory. Asked it to "Do a professional analysis of Tesla stock " and it did ~2wks of professional-level work in ~1hr!

As cool as the new Sora is, gymnastics is still very much the Turing test for AI video. 1/4

Elon opens up about H-1Bs on Nikhil's pod: "Some of the outsourcing companies have gamed the system on the H-1B front, and we should stop the gaming of the system.. but I'm NOT of the school of thought that we should shut down the H-1B program.. that would actually be very bad"

i generated this entire 45s movie clip (audio + video) with claude code + seedance 2 api there's still telltale AI smell, but we should be at full length movies indistinguishable from real ones by the end of the year (veo 5)

Google silently launched a way to describe a speaker's voice "style" in plain English. I've never seen anything like it. It seamlessly does different accents, dialects, tone, and languages. Check out Chinese vs Jamaican and Tamil/English vs Punjabi/English in this clip below!

Lalit Modi, the fugitive founder of the Indian Premier League (IPL), openly admits to.. Hiring a guy to get him a 1550/1600 in the SAT in 1983, faking all his school grades from class 2 to 12 to be 90%+ and eventually went to Duke for his undergrad.

Hanumankind's latest music video is #4 on global Youtube music. He is Sooraj Cherukat, an Indian from Houston who moved to India, worked at Goldman and launched his rap career in his late 20s. Netravalkar juggles his Oracle job with cricket. This is the modern Indian way.

It's insane to me that despite all the tech progress, you can't find a single half-decent place on the internet to read epics of history like the Mahabharata and Ramayana. Until now. All 10,000 pages. In readable modern English. Free. Read it on grand old books.:

Wow, Sonnet 3.7 w/ Thinking just solved a problem no other model could solve yet. "Can you write the most intricate cloth simulation in p5js" Grok 3 and o1 pro had no usable results. This is truly the best "vibe coding" model.

We just solved text-to-speech AI. This model can simulate perfect emotion, screaming and show genuine alarm. — clearly beats 11 labs and Sesame — it’s only 1.6B params — streams realtime on 1 GPU — made by a 1.5 person team in Korea!! It's called Dia by Nari Labs.

I just "vibecoded" a Chess master (~2250 ELO) from scratch that runs locally on a Mac in Rust. I used to play chess semi-competitively, and I'm flabbergasted that you can just speak a 98% percentile chess engine into existence.

Using light as a neural network, as this viral video depicts, is actually closer than you think. In 5-10yrs, we could have matrix multiplications in constant time O(1) with 95% less energy. This is the next era of Moore's Law. Let's talk about Silicon Photonics... 1/9

Billionaire Michael Milken joked “if a US company replaces the US-born CEO with a CEO born in India, I buy the stock” But he reveals he hasn’t backtested the idea. So we did. In the last 15yrs, that would’ve 50x’d your money: 7.5x more $$ and >2x IRR vs S&P500: 30% vs 14%!

Two 23 year old Indians just dropped the #2 open-weight AI voice model in the world, trained purely on free credits! Maya1 is #20 globally, better than even Google's best. 3B params, runs on one GPU and does 20+ emotions with < 100ms latency You can just do things.

Bytedance just dropped realtime voice translation 3x faster than before, with only a ~3s lag! Seed LiveInterp 2 is a full duplex speech-to-speech model with >70% correctness. When this makes it to video calls, it'll open up previously impossible connections.

Sonnet 3.7 generated this entire "AI agent startup" website from scratch. It's superpower is its 64k output token length, which makes the resulting Artifacts way richer. This is better than most actual startup websites.