作为低维生物应该给高维智慧让路了
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
1
4
248
Rare to hear people talking about the largest production training runs on earth. Excited to hear from the goat Rohan Anil at #reverie2026
Bit more than a decade in training, no pain no train. 💪
@BEBischof nerd sniped me into this talk with offer to have a Cat to give the keynote.
Still Training: A Decade of Learning the Hard Way @ Reverie by LanceDB on Nov 5 in SF 🌉
You get to hear me about distributed training from the trenches:
Hardware codesign, async and sync training -- when we almost decided to put async into chips
From big batch gradient descent with parallel boosting from stochastic optimization with duals to online learning where we had regrets, to higher order methods that scale that was almost second order that nobody believed it.
2
5
227
Lei Xu retweeted
Who gets to shape the future of AI?
Today we're bringing together some of the biggest voices in open models, data, and agents to answer that question, including @rbren_dev of @OpenHandsDev, @zlxi02 of @inferact, @GerryFPatia of @boltdotnew, @vinaygo of @Amplitude_HQ and @latkins of @arcee_ai.
The Open Intelligence Summit, co-hosted with @nvidia.
Watch live, starting at 1 PM PT:
Keynote: youtube.com/live/IhWSgZlVizQ
Build: youtube.com/live/eN0TqmeBNhw
Serve: youtube.com/live/fkT-MhcAm-4
Run: youtube.com/live/QMrY1xZ-xGg
Lei Xu retweeted
The best kind of work almost always comes from collaborations!
This time @loldedxd helped me figure out the internal workings on funes. While I knew how to make funes work, he made sure I understand how things run on my system.
It is a fun 30 mins read.
huggingface.co/blog/ariG2349…
Just merged into Lance: a completely new FTS index layout
🚀 9.35x faster index builds (34s → 3.6s)
🚀 8.87x smaller index size (955MB → 107MB)
🚀 3.41x faster phrase queries (21ms → 6ms)
Block bitpacking + delta encoding + new phrase query algo.
Opt in today with LANCE_FTS_FORMAT_VERSION=2
github.com/lance-format/lanc…
We can just build.
I build this as a toy project but now it’s available to all duckdb users over the world. So amazing!
Let me what we should build next 🙌
Excited to announce that Lance is now a 𝗰𝗼𝗿𝗲 𝗲𝘅𝘁𝗲𝗻𝘀𝗶𝗼𝗻 of @duckdb! 🦆🚀
The Lance core extension allows DuckDB users to read and write Lance tables, enabling advanced operations directly within DuckDB, including vector similarity search, full-text search, and hybrid search across local or S3-based datasets.
Lei Xu retweeted
Replying to @lancedb @huggingface
vector search on a 10TB dataset on HF without any separate serving system is crazy !
1/5 Large multimodal blobs don’t have to break dataset workflows.
Images and videos are often treated as external files, separate from metadata and indexes. Once datasets get large, that split makes exploration, curation, and training painful.
Lance changes that on the 🤗 @huggingface Hub. 🧵👇
Lei Xu retweeted
Capping off a cold, wintry, locked-in day ☕️ with this AWESOME PR on a graph benchmark, it totally made my day.
The #Lance format is something to hold close to the heart 🤗, and there's LOTS more to do on the lance-graph front, stay tuned!
See: github.com/prrao87/graph-ben…
Lei Xu retweeted
exa-d uses
@raydistributed pipeline execution for incremental recomputation
@lancedb for surgical column patching per fragment
What does it take to store the web as a database?
exa-d is our internal data framework that orchestrates declarative typed dependencies, sparse updates with precise granularity, efficient and parallel execution across scaling compute, and more.
exa.ai/blog/exa-d
Exa team is amazing early adopter of the @lancedb Lance format. Proudly collaborating with them along the way.
What does it take to store the web as a database?
exa-d is our internal data framework that orchestrates declarative typed dependencies, sparse updates with precise granularity, efficient and parallel execution across scaling compute, and more.
exa.ai/blog/exa-d
3
385
This was the most exciting development during last Christmas.
149
1/5 In this blog post, we discuss what we mean by “multimodal” and the different lenses through which it can be viewed. Rather than thinking of “multimodal” as just a content label (text, images or video/audio), it makes sense to think of it at a *system* level.
Lei Xu retweeted
Alright, let's go! Once the @duckdb team approves my extension bump PR, we'll be able to use all those fancy features!
























