gallen

Founder @Datasent_x• Lossless tokens that preserve structure + compute + governance. Ending the compression vs. structure trade-off for AI infrastructure.

[gallen @gallen1128]

Links:

Models + everything: https://t.co/FrqIdxDy3v

Demo (run your own image): https://t.co/QBaxHxHwmQ

RAG demo: https://t.co/FX0jD3E19Z

[gallen @gallen1128]

Ran end-to-end benchmarks to see how our encoding holds up fed to vLLM at serving scale. Better than expected.

Datasent's tokenizer makes the smallest representation of a piece of data that still retains what downstream workloads need. Here, compressing the visual tokens going into a vision-language model to a fraction of their length, plus a tiny adapter that teaches the model to read the compressed version.

On Qwen2-VL-7B:

→ 3-4x more requests per GPU (almost a quarter of the hardware for the same volume)
→ accuracy went UP, +8 points
→ ~5-7x smaller on disk
→ time-to-first-token 2-4x faster

Why: shorter visual prefill = less compute per request, and prefill attention scales with the square of the sequence, so each GPU serves a lot more. Compounds when you encode a catalog once and query it many times, which is what most multimodal RAG already does.

All on Hugging Face: model cards, demos you can run on your own images, a RAG demo. Link below.

[gallen @gallen1128]

I really like, “fewer tokens wasted, not fewer tokens used.” Although, not every token costs the same.

Everything on this list optimizes how many tokens move through the system. But once you’re multimodal, the cost stops being about count. A single image is hundreds to thousands of tokens, attention is quadratic in sequence length, and the visual side quietly dominates the bill, untouched by caching or routing. So the obvious direction after this isn’t fewer tokens, it’s cheaper tokens, which is a representation and compression question rather than an orchestration one. Different axis, stacks on top of all of the above

[gallen @gallen1128]

We started Datasent because I was tired of the same false choice in every data system I built:

Compressing meant losing structure, but preserving structure exploded costs, and to make it ML/LLM-ready we had to add another lossy pipeline.

So we created a new mathematical primitive that ends the trade-off.

It’s called Datasent - and the full research suite just dropped.

[Polymesh @PolymeshNetwork]

ICYMI 👉🏼 Exciting news last week as the Polymesh Association and BDACS signed a strategic partnership agreement supporting the discovery and development of promising #RWA projects.

This agreement will bolster Polymesh’s reputation as a world-class infrastructure for #tokenization in South Korea, where it is already well-known for having been proposed as the RWA technology model for the #Busan Digital Asset Exchange.

Together Polymesh and BDACS will encourage the blockchain industry towards a regulatory-compliant #DigitalAssets ecosystem through collaborations expanding Polymesh’s RWA and tokenized securities protocol in South Korea.

The partnership synergy will also support the city of Busan’s development into a global #BlockchainHub by jointly cooperating to activate the Busan Blockchain Special Zone through the tokenization and listing of various promising RWA products.

https://t.co/q983BglpWO

#SouthKorea #KoreaBlockchain #BWB2024 #BlockchainPartnership

[Polymesh @PolymeshNetwork]

We’re coming for #1

[gallen @gallen1128]

https://t.co/5x9zmJAle9

[Datasent @datasent_x]

Check out my latest article: Unlocking the Future of Data Analytics with Innovative Technology https://t.co/OLUirtOffR via @LinkedIn

[gallen @gallen1128]

@mrajchan Enjoyed this post. Are these companies going to be residing outside capital?

[gallen @gallen1128]

@connortomkies_ Awesome!

[gallen @gallen1128]

@clifflerner @getsaturdayapp @DarrellLerner Hmm, very interesting!

[Boris Berenberg @BorisBerenberg]

Spent a couple hours chatting with @sirbiro about founding and selling my company

https://t.co/hSF0G2iJCu

[Nikhil Goel @ngoel]

UPDATE: Utilizing learnings from the previous 4 years of flight testing, our Midnight aircraft recently finished phase 1 of its flight test program in ~3 months. This is significantly faster than @flyarcher's full-scale prototype aircraft, Maker, and is an example of how our team continues to accelerate toward being production-ready.

Get the full story here: https://t.co/mYy9Bt3tbz

[Polymesh @PolymeshNetwork]

Reminder to submit your questions! 👇🏼👇🏼👇🏼

[Polymesh @PolymeshNetwork]

Daily #tokenization fact ☕

The path to mass adoption of #SecurityTokens includes overcoming regulatory hurdles ⚖ 🏛 and educating market participants about their potential benefits 📖 📝

We're doing just that – share this tweet to let the world know 🐦 🔃

[gallen @gallen1128]

#BinancePOLYX

[Binance @binance]

Which project has the strongest community?

Support your favourite protocol to share $10,000!

To enter:
🔸 Retweet & follow @binance
🔸 Share your project's hashtag (e.g. #BinanceBTC)
🔸 Tell us why you chose that project

Rules and T&Cs ➡️ https://t.co/j8miB6wb8h

[gallen @gallen1128]

@PolymeshNetwork Polyx, a product of linear Factors.

[gallen @gallen1128]

https://t.co/4dUQTY16KT

[gallen @gallen1128]

Math... http://t.co/lvbwDWjG