Robert Krajewski (@rob_krajewski) profile picture

2x Founder | Building in San Francisco | ML PhD Dropout

San Francisco, CA
Joined May 2014
Photo by Robert Krajewski (@rob_krajewski)
This is how frontier AI labs swallow memory startups. Every recalled memory is a data point for RL on the next model. All startups tappinh into LLM context will sooner or later be wiped out. With the recent shorter model release cycle, likely sooner.
Memory is what allows Claude to learn from experience. In a new Engineering Blog post, we share how “dreaming” helps Claude organize what it’s learned, correct its mistakes, and find what it needs more easily the next time. anthropic.com/research/compo…
1
1
122
I build heappi - the whole Pi harness, including codemode, running in snapshotable QuickJS heap. github.com/topos-sh/heappi Large portion of harness'es job is waiting for user messages / LLM Streams / tools etc. Heappi snapshots Pi on any await, including awaits in the code mode
1
1
1
3
189
While snapshotted, heappi can wait seconds or days for next user message or tool result without occupying RAM. It resumes in 4 to 25 ms without replaying the whole program. The heap is stored incrementally in SQLite / S3 so program can be resumed anywhere (given the same build)
1
2
22
Project is inspired by: * @tobi's tweet about github.com/stencil-hq/dquick… * recent release of github.com/hydro-project/inf… * Pi durable and Pi's code mode: github.com/earendil-works/pi… * a little bit of github.com/denoland/celld for storing heap incrementally.
2
39
New W27 RFS from YC. Two themes dominate: 1. "Own your own inference / training", and 2. "Supply data centers / robots with components".
The bottlenecks in AI are moving beyond the models. We’re excited about startups building the hardware, infra, financial systems, and specialized software to bring AI into the real economy, from factories and data centers to government and space. ycombinator.com/rfs
2
40
Robert Krajewski retweeted
The bottlenecks in AI are moving beyond the models. We’re excited about startups building the hardware, infra, financial systems, and specialized software to bring AI into the real economy, from factories and data centers to government and space. ycombinator.com/rfs
161
259
161
2,563
589,324
70% of internet runs on PHP. This is going to be interesting.
Effective Immediately, all organizations using PHP for their systems are advised to be prepared for an update addressing two critical vulnerabilities. PHP is widely used for web applications. These vulnerabilities often provide a direct path to the server's file system or database. If an immediate update is not feasible, it is recommend that organizations take the system offline until it can be fully mitigated. NSA will follow up shortly with more specific guidance.
2
33
OpenAI strikes again. This time with 722 math proofs or counterexamples including Quasi-Riemann hypothesis, and large piece of Hodge conjecture.
We’re releasing a broad range of new mathematical results produced by an internal frontier model. We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results. github.com/openai/math
1
95
Robert Krajewski retweeted
We’re releasing a broad range of new mathematical results produced by an internal frontier model. We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and public recommendations to inform how we release these results. github.com/openai/math
1,896
6,272
4,764
40,544
22,846,470
Another growing AI niche. Sandboxes for agents: Cloudflare Sandboxes (cloudflare.com/products/sand…) Modal (modal.com/products/sandboxes) Fly io sprites (fly.io/sprites/) Vercel Sandbox (vercel.com/docs/sandbox) Daytona (daytona.io/) E2B (e2b.dev/) Blaxel (blaxel.ai/platform/sandboxes) Runloop (runloop.ai/) Microsandbox (microsandbox.dev/) Box (box.ascii.dev/) Superserve (superserve.ai/) Freestyle (freestyle.sh/) Northflank (northflank.com/) Novita (novita.ai/sandbox) They compete on SDK lovability and price. Large incumbents like Cloudflare and Vercel add value by integration with their ecosystems.
2
47
Mathematicians were committing their whole lives to solving open problems and learning along the way. AI is taking away the main human driver, being the first, and hence the letter (mathandai.org/). But the assumption that math needs humans, may soon no longer be true, however it sounds.
2
23
Multi-harness orchestrators / coding tools: Zellij (zellij.dev/) Herdr (herdr.dev/) cmux (cmux.com/) Zeron (zeron.sh/) TermRover (termrover.sh/) moshi (getmoshi.app/) AO (useao.dev/) Traycer (traycer.ai/) Synara (trysynara.com/) Orca (onorca.dev/) Paseo (paseo.sh/) Emdash (emdash.com/) Superset (superset.sh/) super engineering (super.engineering/) Superconductor (superconductor.com/) Exponential (exponential.at/) Helmor (helmor.ai/) At this rate, there will soon be one orchestrator company for every non-orchestrator company.
1
3
136
New (unpublished) model from OpenAI solved 200yo Millennium Prize Problem in math. This problem was so important, that a $1M award was offered by Clay institute. Solution and derivatives will have direct impact on understanding fluid mechanics. Pace of progress in AI unprecedented. True beginning of infinity!
We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics. The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra. The problem concerns whether the description of smooth three-dimensional fluid motion modeled by the Navier-Stokes equations can break down. It has remained unresolved for roughly 90 years.
1
26
Unimaginable. A few years ago it was completely unpredictable. People were preparing for years of hard work by hundreds of mathematicians. We were born too early
Checking that a major mathematical proof is correct can take years. Formalization—converting the mathematical reasoning into a form computer proof assistants like Lean can verify—can help. Last month, Claude completed the first formalized proof of Fermat’s Last Theorem, one of the most famous theorems of all time. This was a project experts thought would take many years. It is the largest Lean proof ever written. Fermat’s Last Theorem was first proven in 1995 by Sir Andrew Wiles, more than 350 years after it was conjectured. Our proof, which totals over 13 million lines of code, provides machine verification. More importantly, it proves over 29,000 other theorems that the proof requires, across many areas of math which had never before been formalized. We see this as a major step in the long process of firming up the core of mathematical knowledge, building on work from three centuries of mathematicians and hundreds of contributors to Lean and Mathlib. We are optimistic that AI-assisted verification of mathematical proofs will help reduce the burden of refereeing mathematics in an era where more proofs are being produced than ever before. You can read about the process on our Science Blog: anthropic.com/research/forma… And see the complete proof on GitHub: github.com/anthropics/fermat…
3
2
31
YC's Fall Request for Startups dropped one week before the application deadline. "Multiplayer AI" and "Self-Maintaining APIs" are two sides of the same problem: keeping multiple distributed agents up to date, within and across organizations.
2
3
70
Yesterday's @vllm_project x @CrusoeAI meetup at @a16z was a perfect refresher on what's up in the inference layer. Thanks for hosting and sharing!
3
64