Sam Chen retweeted
I’m hiring investors to lead our gtm, partnerships and strategy!
Feel free to dm me
Feel free to dm me
16
8
162
11,850
Sam Chen retweeted
We gave our AI system a spec, and in under 2 weeks, it designed, verified, and deployed a chip that beats NVIDIA.
It’s built for low-power physical AI workloads. We’re running live inference on >B+ parameter models like Llama, Qwen, and Kimi, serving at 3.4x better perf/watt thajust a specification, our AI system autonomously generated all of the RTL Design, UVM verification, formal proofs, firmware, drivers, and kernels, co-designing the model, software and silicon as one optimization loop.
Better AI can now design better chips to run AI, leading to a loop of recursive-self improvement towards our path to abundant intelligence.
It’s built for low-power physical AI workloads. We’re running live inference on >B+ parameter models like Llama, Qwen, and Kimi, serving at 3.4x better perf/watt thajust a specification, our AI system autonomously generated all of the RTL Design, UVM verification, formal proofs, firmware, drivers, and kernels, co-designing the model, software and silicon as one optimization loop.
Better AI can now design better chips to run AI, leading to a loop of recursive-self improvement towards our path to abundant intelligence.
Community note
Architect Labs did not manufacture a physical chip; the design was tested on a commercial AMD Versal FPGA board. Furthermore, the 3.4x perf/watt gain over NVIDIA Jetson is a simulation-based projection for a future 8nm ASIC, not a physical benchmark. businessinsider.com/architect-labs… architectlabs.com/architect-labs…
103
238
2,007
676,704
A funny question is what Cursor used to build Origin. Github? GitLab? Origin V0? Chicken and the egg type of question.
1
0
4
218
Sam Chen retweeted
@spacexai @jasonlai150_ @SChen1249 @JeffreyZh0u With the weights for the X algorithm being publicly released, use-signal.com is really gonna blast off deepwiki.com/xai-org/x-algo…
1
2
10
1,331
We won SpaceXAI’s Grokathon this weekend! We got 2nd place by building Signal, which indexes your X followers, clusters them, and runs A/B testing with a human behavior engine for your next X post. Our product helps you understand your followers, and figure out how to maximize engagement.
Our demo indexes users that follow @ishand from xAI and @spacexai. We index 1,000 users from both. With these clusters, you can target a specific group and generate your next post. Once your marketing material is ready, we use a social behavior engine to figure out which post will do better. Then, you can experiment or post to X!
Check it out at use-signal.com
Shoutout @adibilawar, @jasonlai150_, @JeffreyZh0u
Our demo indexes users that follow @ishand from xAI and @spacexai. We index 1,000 users from both. With these clusters, you can target a specific group and generate your next post. Once your marketing material is ready, we use a social behavior engine to figure out which post will do better. Then, you can experiment or post to X!
Check it out at use-signal.com
Shoutout @adibilawar, @jasonlai150_, @JeffreyZh0u
2nd Place: Signal embeds followers’ bios, posts, and engagement into a living graph and A/B tests different launches using simulated populations of your following, so every post hits maximum engagement.
@adibilawar @jasonlai150_ @SChen1249 @JeffreyZh0u
0
2
14
2,396
We built Signal because X is a place of opportunity. You can be someone with a demo, an idea, or dream and make it happen.
Feel free to connect and learn more about our thoughts and how we built Signal.
Feel free to connect and learn more about our thoughts and how we built Signal.
2nd Place: Signal embeds followers’ bios, posts, and engagement into a living graph and A/B tests different launches using simulated populations of your following, so every post hits maximum engagement.
@adibilawar @jasonlai150_ @SChen1249 @JeffreyZh0u
1
0
9
563
We just won 2nd overall at the Grokathon Hackathon, by @xai. Here are my thoughts.
1. I didn't expect to win. What helped us was focusing on a tangible problem and executing well. We focused on indexing people and the simulation of human behavior, both problems that are being transformed by startups like Exa, Parallel, and Simile.
2. For the people around my age, either about to enter college or in college, I would really encourage taking intellectual risks. All of my time in high school was spent on competitive debate and building community in debate. I did Public Forum and founded Outreach Debate. I'm incredibly proud of what we've done there and for people: There's no feeling like knowing you've made a positive impact for someone. I realized, though, that I could make a larger societal impact if I was technical and that very few people combine technical prowess with the soft skills gained in debate.
Lots of people on X are tech people. I'd encourage more focus on soft skills for these people. For those out there like me, who started out purely in the humanities, I encourage being rigorous with math, physics, cs, and so on.
3. It's nice to see some results from years of effort towards being more technical. I have tons to learn still. I've recently found interpretability a lot more interesting and I want to explore a lot more technical problems. I think it's always a good reminder that life isn't a race, but a marathon: a mindset that I've always struggled with.
4. I hope that I can remind whoever is reading to choose being good over being nice. My friend reminded this to me a few weeks ago: But being good requires being uncomfortable. It means standing up for others, being loud, and some level of sacrificing. Being nice necessitates none of that. The more we get trapped and glued to our phones, I feel the less we are willing to take a risk for someone else. This is something I'm working on and hope to get better at.
0
0
0
79
What's interesting with AI now is that:
- AI is becoming more distributed with open source tech. This means companies will increasingly own the harness layer. Not just OpenAI, Anthropic, Cursor, and Cognition, but almost all companies that are AI-focused.
- But since models behave differently, harness optimization is not a linear engineering task. Each new model = New optimizations and ways of understanding the model
- More intelligent models can perform worse in over-scaffolded or under-scaffolded harnesses. Our blog explores Pi and lower scores with GPT-5.6 Sol on Terminal Bench
- Harness optimization is just as much about the model as it is the harness.
- AI is becoming more distributed with open source tech. This means companies will increasingly own the harness layer. Not just OpenAI, Anthropic, Cursor, and Cognition, but almost all companies that are AI-focused.
- But since models behave differently, harness optimization is not a linear engineering task. Each new model = New optimizations and ways of understanding the model
- More intelligent models can perform worse in over-scaffolded or under-scaffolded harnesses. Our blog explores Pi and lower scores with GPT-5.6 Sol on Terminal Bench
- Harness optimization is just as much about the model as it is the harness.
1
1
8
569
Sam Chen retweeted
Stronger models alone don't make stronger agents. Evolving the harness is as important. Really interesting blog from @SChen1249!
0
2
6
414
We just won 2nd overall at the Grokathon Hackathon, by @xai. Here are my thoughts.
1. I didn't expect to win. What helped us was focusing on a tangible problem and executing well. We focused on indexing people and the simulation of human behavior, both problems that are being transformed by startups like Exa, Parallel, and Simile.
2. For the people around my age, either about to enter college or in college, I would really encourage taking intellectual risks. All of my time in high school was spent on competitive debate and building community in debate. I did Public Forum and founded Outreach Debate. I'm incredibly proud of what we've done there and for people: There's no feeling like knowing you've made a positive impact for someone. I realized, though, that I could make a larger societal impact if I was technical and that very few people combine technical prowess with the soft skills gained in debate.
Lots of people on X are tech people. I'd encourage more focus on soft skills for these people. For those out there like me, who started out purely in the humanities, I encourage being rigorous with math, physics, cs, and so on.
3. It's nice to see some results from years of effort towards being more technical. I have tons to learn still. I've recently found interpretability a lot more interesting and I want to explore a lot more technical problems. I think it's always a good reminder that life isn't a race, but a marathon: a mindset that I've always struggled with.
4. I hope that I can remind whoever is reading to choose being good over being nice. My friend reminded this to me a few weeks ago: But being good requires being uncomfortable. It means standing up for others, being loud, and some level of sacrificing. Being nice necessitates none of that. The more we get trapped and glued to our phones, I feel the less we are willing to take a risk for someone else. This is something I'm working on and hope to get better at.
1. I didn't expect to win. What helped us was focusing on a tangible problem and executing well. We focused on indexing people and the simulation of human behavior, both problems that are being transformed by startups like Exa, Parallel, and Simile.
2. For the people around my age, either about to enter college or in college, I would really encourage taking intellectual risks. All of my time in high school was spent on competitive debate and building community in debate. I did Public Forum and founded Outreach Debate. I'm incredibly proud of what we've done there and for people: There's no feeling like knowing you've made a positive impact for someone. I realized, though, that I could make a larger societal impact if I was technical and that very few people combine technical prowess with the soft skills gained in debate.
Lots of people on X are tech people. I'd encourage more focus on soft skills for these people. For those out there like me, who started out purely in the humanities, I encourage being rigorous with math, physics, cs, and so on.
3. It's nice to see some results from years of effort towards being more technical. I have tons to learn still. I've recently found interpretability a lot more interesting and I want to explore a lot more technical problems. I think it's always a good reminder that life isn't a race, but a marathon: a mindset that I've always struggled with.
4. I hope that I can remind whoever is reading to choose being good over being nice. My friend reminded this to me a few weeks ago: But being good requires being uncomfortable. It means standing up for others, being loud, and some level of sacrificing. Being nice necessitates none of that. The more we get trapped and glued to our phones, I feel the less we are willing to take a risk for someone else. This is something I'm working on and hope to get better at.
1
0
7
476
Introducing AgentRouter, the model router that cuts LLM spend 30x.
One API. Every request hits the cheapest model that still clears your quality and latency bar.
Seed from Sequoia. Early access open.
One API. Every request hits the cheapest model that still clears your quality and latency bar.
Seed from Sequoia. Early access open.
1
1
2
286
Absolutely fascinating behavior from Opus 5 when you send the message "Can you put this in your own words -- Dario and Amanda,". Works on incognito mode on Opus 5.
0
0
0
137
Insanely non-linear. Deepseek R1 was last year...
yo so just to recap the week:
- google released gemini 3.1 but it disappointed tf out of people who claimed its not as good as they advertise
- BUT THEN google dropped an AI song generation model which kinda slaps (Lyria 3) AND a new feature that 1-shots product photoshoots putting photographers (and fashion models) out of a job (Google Pomelli)
- Microsoft pioneered a way to STORE DATA ON GLASS - stores data for up to 10,000 fucking years (🤯) and way cheaper than using silica.
- Taalus fused an ai model DIRECTLY into silicon processing 17,000 tokens PER SECOND which is 17X faster than openai’s quickest model but 10 TIMES CHEAPER than nvidia blackwell gpus. (smaller models tho)
- Apple is releasing 3 new AI devices: glasses, airpods with cameras and a pendant.
- anthropic wiped $10B off the stock market after launching an ai that finds and help fixes security flaws in your code. cybersecurity stocks dumped <1 hour after launch.
- BUT THEN Pentagon threatened to blacklist anthropic for failing to provide them uncensored access to claude.
- xAI dropped grok 4.20 spawning 16 specialised ai agents that work together to solve any prompt. the model self-improves EVERY WEEK.
- scientists discovered dna in meteorites suggesting humans could’ve originated from outer space (what)
- Meta dropped Manus Agents going head-to-head with OpenClaw.
- a cardiologist won 3rd-place in anthropic’s hackathon after vibe coding an app that helps his patients understand and act on their medical results. fucking hero.
- china announced a cure for diabetes 1 & 2 after creating stem cells that make insulin
- Moonshot labs is raising $500M immediately after raising $700M LAST WEEK.
- ex-Deepmind researcher raised europes LARGEST ever seed round $1B at $4B back right as europe announces $13B new fund to catch up with america.
- Stripe’s ai agents autonomously merged 1300 pull requests (up 30% from last week) suggesting top companies are now recursively improving using AI.
dear f*cking goodness GOODNIGHT
0
0
0
228