Building an AI doctor @superpower. @thielfellowship @mozillabuilders
Joined April 2020
- Tweets424
- Following51
- Followers1.1K
- Likes177
Kenta Iwasaki retweeted
the dumbest smart people you know are debating temperature and rngs and skewing probability distributions over semantically similar "words" because they heard about word2vec once and believe "word choice" has no functional purpose beyond a canonical dictionary definition and a cosine similarity in latent space
"isn't hiding a signature in text the same as hiding it in rgb in a pixelated image?", cries the clever-by-half hot-taker
Kenta Iwasaki retweeted
hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend fable for working during the world cup final
((1+xy)^3 z + y^2 (1+xy) (4+3xy), y + 3 x (1+xy)^2 z + 3 x y^2 (4+3xy), 2 x - 3 x^2 y - x^3 z): \C^3\to \C^3, has jacobian determinant -2, and sends (0, 0, -1/4), (1, -3/2, 13/2), and (-1, 3/2, 13/2) to (-1/4, 0, 0)
Kenta Iwasaki retweeted
Legends don't guess. They measure.
Giannis Antetokounmpo is our first brand ambassador. 6.30.26
Kenta Iwasaki retweeted
Andrew's interview with JetBrains
youtube.com/watch?v=iqddnwKF…
Open-sourced quickjs-bun: lightweight, fully-featured, idiomatic `bun:ffi` bindings to QuickJS. Wrote it at @superpower for giving our AI agents a (really) fast sandboxed JavaScript interpreter for tool calling (what some call "codemode"). github.com/superpowerdotcom/…
Kenta Iwasaki retweeted
I'm usually not one to write thought pieces without much technical depth. But here we go.
Slow the fuck down.
mariozechner.at/posts/2026-0…
Can't believe I got to build this with such a cracked team. Honored to be a part of @superpower ❤️.
Kenta Iwasaki retweeted
Postgres as data platform: make heterogeneity a first-class engine property—SLA-aware planning, pluggable execution, and machine-readable feedback (visibility, degradations, sources). Otherwise it's just storage+glue. noodata.substack.com/p/postg…
#PostgreSQL
Made a PR to add support for Anthropic `search_result` blocks and custom content documents to @aisdk. Citations are fully supported.
cc @aayushkapoor_ @lgrammel github.com/vercel/ai/pull/11…
Given that text has very little information and is non differentiable, wouldn’t it make more sense to produce models off of differentiable features such as images and audio instead? We could still work off of existing text data by rendering them into images.
Open-sourced my notes for computing log-Mel spectrograms on CPUs 129x faster. OpenMP, thread-local scratch buffers, hand-rolled real DFT's, sparse CSR-like Mel dot-products, and @bunjavascript FFI. ≤ 1e-3 max abs error vs. OpenAI Whisper/PyTorch/librosa. github.com/lithdew/log-mel-s…
Update: made it 521x faster and got the max abs error down to ≤ 1e-4 by implementing a specialized mixed-radix FFT. Code is in repo.
Published native, up-to-date chDB bindings for @bunjavascript. chDB is automatically downloaded upon install, resources are properly cleaned up when garbage-collected, and streaming queries and chDB's connection-based API's are supported. github.com/lithdew/chdb-bun
Given 1 writer/N reader threads on SQLite 3.49.2 in WAL mode, there are cases where reader threads may acquire an EXCLUSIVE-WRITE lock while reading from the WAL which could then cause the writer thread to fail to acquire an EXCLUSIVE-WRITE lock and yield the error SQLITE_BUSY.
Went through this rabbit hole debugging for 6 hours b/c I was getting SQLITE_BUSY errors while writing a multithreaded SQLite worker pool with Bun.
Had to disable WAL auto-checkpointing, read WAL/pager traces, play with compiler defines, etc.
Ty @jarredsumner for all the help 🙏.
