brandon wang

... | prev undergrad @miteecs and @mitbiology, @cartesia @janestreetgroup

[brandon wang @fluorane]

happy to announce that we've gotten rid of tokenizers!

especially excited with what we've replaced them with: end-to-end trainable modules that not only learn to group characters into (sub)words, but can iterate to group words into phrases and further higher-order concepts

see @sukjun_hwang's thread for more details 👇

[Sukjun (June) Hwang @sukjun_hwang]

Tokenization has been the final barrier to truly end-to-end language models.

We developed the H-Net: a hierarchical network that replaces tokenization with a dynamic chunking process directly inside the model, automatically discovering and operating over meaningful units of data

[brandon wang @fluorane]

so what the hell did openai put into gpt image 2

imo still the most impressive model of the year so far

[brandon wang @fluorane]

surely managing a team of agents is also ddr

[Sergiy Galyonkin @galyonkin]

A conversation I had earlier today.

[Harry Heymann 🥑 @harryh]

I really miss the "information wants to be free" left of the 90s / early 2000s. We have to go back.

[brandon wang @fluorane]

@frank_liquid idt its problem difficulty, its more like if you have more leverage you should work harder (eg rearrangement ineq)

[brandon wang @fluorane]

@_isomerism weekend is visible on the chart, feeling bearish

[brandon wang @fluorane]

@suekhim congrats! looks very cool

[brandon wang @fluorane]

@KabirGoel @chongz be the change you wish to see in the world

[brandon wang @fluorane]

@elipughresearch @radiuskia it will continue to be posted until agi is achieved

[brandon wang @fluorane]

happened again (oc @radiuskia)

[Cartesia @cartesia]

Cartesia Ink-2 debuts as #1 for accuracy on the brand-new streaming speech-to-text leaderboard from @ArtificialAnlys! We designed Ink-2 from the ground up for voice agents - with low latency, eager transcripts, and semantic endpointing.

[brandon wang @fluorane]

kimi agent swarm has been achieved domestically, part 2

[Claude @claudeai]

Also new in Claude Code: dynamic workflows (research preview).

For the hardest tasks, Claude makes a plan, runs hundreds of parallel subagents, and verifies its work before reporting back. Think a migration touching hundreds of files.

Read more: https://t.co/7gt06kGkDN

[brandon wang @fluorane]

@elipughresearch of course you put tadej into the eval set ...

[brandon wang @fluorane]

@saurabh_shah2 yeah! should be fun :D

[brandon wang @fluorane]

@saurabh_shah2 i think it really likes adding unit tests around error types for some reason?

[brandon wang @fluorane]

@saurabh_shah2 codex gave me this nice function with example call self._to_device_tensor("x", x)

[brandon wang @fluorane]

@chongz @zhennydez its true, being homeless is better than having no gpus

[brandon wang @fluorane]

@0xpiapark @pierrecomputer congrats pia! very hype, good luck :D

[brandon wang @fluorane]

@serenaa_ge very cool benchmark, congrats! failure modes analysis is really interesting

[brandon wang @fluorane]

@frank_liquid congrats! very smooth experience

[brandon wang @fluorane]

@Alan_Ma_ huawei duv is cooking indeed

[brandon wang @fluorane]

the dim sum restaurant im at is showing a semiconductor lithography report from a hong kong tv channel, amazing