Tokenizers v1 RC is out.
Same token IDs as v0.23, 3–30× faster encode, ~76% linear scaling on 8 cores.
Main changes: SIMD bitstream pre-tokenization (bitcannon) instead of regex, no-alloc merge in a scratch buffer, thread-local word cache, smaller crate.
cargo add tokenizers --pre
huggingface.co/blog/tokenize…
Kudos @art_zucker @SimonBrandeis @LucSGeorges @LysandreJik
There’s a lot to like about @antigravity cli
It has all the hot staff like remote control and /voice
But I really like the TUI — rich diffs, artifacts, diagrams, /agents and /tasks panels. Feels intentionally designed, not a chat transcript poured into ANSI
Only problem Google restricts it to Gemini
Recently @shengzheyao exposed GOOGLE_GEMINI_BASE_URL. It speaks the older Gemini `generateContent` format, so I added that protocol to Swobu.
Now I can run unpatched w/ local Qwen + OpenAI for web-search fallback.
Local LLM doesn't have to mean a local UI.
I have Qwen running on my machine.
Swobu gives it a public HTTPS endpoint, so I can use the same local model directly from @TypingMindApp in the browser.
No VPN. No provider credentials on the remote side.
@tdinh_me - thought you might enjoy this one :)
Used my local LLM setup from an untrusted phone.
No keys on it.
I changed the backend anyway.
Then I killed the Share.
TLS ends on the owner machine. The relay never sees the Share bearer or the application HTTP in plaintext.
Keys, routing, and provider creds stay there.
swobu.com/blog/https-routes/
RTK reported 349.2M tokens “saved.”
@QuesmaOrg measured +17% average task cost on DeepSeek.
For agents: tokens removed ≠ tokens saved
This generalizes beyond RTK: caching, pruning, routing, even “cheaper” models can save locally while making the full trajectory more expensive.
quesma.com/blog/does-rtk-mak…
Good benchmark from @bartkotrys + @jakozaur
Founders have $10k-$100k AWS/Azure/GCP credits expiring while paying cash for Claude/OpenAI lol
Cloud Creditmaxxing!
Swobu routes AI across clouds/providers anyway. might as well route by expiring credits too
Who has the most ridiculous stranded credit balance? 💸💸
I heard you like Ox
so I packed 4 free Ox Alpha providers:
→Command Code
→OpenCode Zen
→Nous
→OpenRouter
Behind one Swobu endpoint
github.com/swobuforge/swobu
1/ "Language Agents: Foundations, Prospects, and Risks" #emnlp2024
Tutorial by @ysu_nlp @Diyi_Yang @ShunyuYao12 @taoyds
📑 Slides: tinyurl.com/language-agent-t…
🎥 Video: tinyurl.com/agent-tutorial-r…
4/ RAG → HippoRAG (RAG + Knowledge Graph + PageRank) by @bernaaaljg
Outperforms SoTA multi-hop QA by up to 20%
While being 10-30x faster and 6-13x cheaper.
📄 Paper: arxiv.org/abs/2405.14831
🐙 Code: github.com/OSU-NLP-Group/Hip…
"The Little Book of Deep Learning" 140 pp📖
Book by @francoisfleuret 🫡
fleuret.org/public/lbdl.pdf
