Keep on to blur preview images; turn off to show them clearly
RT @SemiAnalysis_: BREAKING CUDA MOAT EXPANDS: Today, NVIDIA has acquired SchedMD, makers of SLURM, a widely used "open source" workload sc…
SemiAnalysis Boutique AI Infrastructure Research and Consulting DMs are open for consulting, quotes, or to talk shop
RT @a1zhang: Super cool work, I'm very excited to see more efforts getting sandboxing / inference for RLMs to work! We're also cooking so…
Asst professor @MIT EECS & CSAIL (@nlp_mit). Author of https://t.co/VgyLxl0oa1 and https://t.co/ZZaSzaRaZ7 (@DSPyOSS). Prev: CS PhD @StanfordNLP. Research @Databricks.
RT @huntermbown: Couldn't get the recursive language model idea out of my head after seeing @JoshPurtell post about @a1zhang and @lateinter…
Asst professor @MIT EECS & CSAIL (@nlp_mit). Author of https://t.co/VgyLxl0oa1 and https://t.co/ZZaSzaRaZ7 (@DSPyOSS). Prev: CS PhD @StanfordNLP. Research @Databricks.
My biggest takeaways from @embirico (OpenAI Codex Product Lead): 1. OpenAI’s initial Codex product was “too far in the future.” It ran in the cloud asynchronously, which was great for power users but hard for newcomers. Growth exploded when they brought it back to where engineers already work: inside their code editor, on their own computer. Codex usage has grown 20x in the past 6 months. 2. OpenAI built the Sora Android app—which hit #1 in the app store—in just a few weeks with two or three engineers, with the help of Codex. The Sora app went from zero to employee testing in 18 days, then launched publicly 10 days later. Codex helped by analyzing the existing iOS app, generating work plans, and implementing features by comparing both platforms simultaneously. 3. The key to getting value from Codex: give it your hardest problems, not your easiest. These tools are built to tackle gnarly bugs and complex tasks, not simple ones. Start with something you’d otherwise spend hours on. 4. Writing code may become the universal way AI accomplishes any task. Rather than clicking through interfaces or building separate integrations, AI performs best when it writes small programs on the fly. This suggests that coding ability should be built into every AI assistant, not just specialized programming tools. 5. Designers at OpenAI now write and ship their own code. The design team maintains a fully functional prototype built with AI assistance. When they have an idea, they code it directly, test it, and often submit it for production themselves. Engineers only step in when the codebase is particularly complex. 6. Even if AI models stopped improving tomorrow, there are still years of product work left to unlock their potential. The technology is ahead of our ability to use it optimally. 7. The biggest bottleneck to AI productivity isn’t the AI; it’s how fast humans can type. The limiting factors are how fast you can type prompts and how quickly you can review AI-generated work. Until AI can validate its own output more reliably and surface help proactively, we won’t see the full productivity gains these tools could deliver. 8. Writing code is becoming less fun than reviewing AI-written code. Engineers love the creative flow of building. Now they’re spending more time reading what the AI produced. The next challenge is making that review process faster and more satisfying. 9. New AI models can now work continuously for 24 to over 60 hours on a single task. A technique called “compaction” lets the AI summarize what it’s learned before running out of memory, then continue working in a fresh session. This enables overnight or multi-day autonomous work that wasn’t previously possible. 10. If you’re starting a company today, deep understanding of a specific customer matters more than being good at building. Building is getting easier. Knowing what to build—and for whom—is the real advantage now.
Deeply researched product, growth, and career advice
RT @USOPM: TUNE IN: OPM Director @skupor will join @tbpn today at 4:30pm ET to hear all about the launch of @USTechForce!
Dir., Office of Personnel Management (previously, MP at a16z); Author of Secrets of Sand Hill Road; father of three amazing/crazy/beautiful girls.
This is an important issue for federal employees planning retirement. Long HR processing times are a real challenge which is exactly why we’re pushing electronic submissions that move cases 2x faster and modernizing outdated processes. OPM is also working to ensure retirees receive interim payments as quickly as possible - 75% are automatically going into automatic interim pay with the rest completed within 7 days of application receipt. https://t.co/YgAdE0Stit
Dir., Office of Personnel Management (previously, MP at a16z); Author of Secrets of Sand Hill Road; father of three amazing/crazy/beautiful girls.