Sourabh

@_Sagiquarius_ @GergelyOrosz @reach_vb Marketing consultants?

@GergelyOrosz @reach_vb I never understood why the machines, especially GPT models, seem to insist on injecting these types of phrases and vernacular. Like they’re trained on such a large corpus of human data, who fucking talks that way?

The amount of AI writing on OpenAI's docs makes me sick. Filler sentences for nothing. Mannerisms and phrases humans would not write. Has a human even read this? And all of this will change how other docs are written, and how we all talk - for the worse IMO. 🤮

this short-video format shouldn’t exist. I already have the memory of a goldfish. It doesn’t need to get any worse

didn't realize model was automatically switched to GPT5.6 Sol

should have taken a nap

The tau Voice results are pretty wild: GPT-5 reasoning got 85% in text, while voice agents were only 31–51% even in clean conditions, and 26–38% in more realistic ones. That gap feels like the real state of voice AI right now.

adding voice makes the whole thing brutally harder. Same smart model, but now add STT errors, accents, noise, interruptions, turn-taking, latency, and the occasional “sorry, can you repeat that?” A great LLM does not automatically become a great voice agent.

Went down the tau bench(#home" target="_blank" rel="nofollow noopener">taubench.com/#home) rabbit hole today and honestly, this is the kind of eval I wish we saw more often. Not “can the model answer a question?”, but “can it actually help a user finish a messy, multi-turn task while following policy and using tools?”

@sumanthd17 you might not need them in the future

Maybe a lesser known fact, but Seoul is one of the better places to get a mechanical keyboard #icml

Recently had to go through this bitter lesson: x.com/Curious_Monkey…

The deeper lesson: tokenization is not just preprocessing. It defines the atomic units of thought available to the model. A model can only reason over the sequence you give it. Change the units, and you change what patterns are easy or hard to learn.

Tokenization-free LLMs? Remember the “count the r’s in strawberry” example that tripped up LLMs? Tokenization is a big reason why. Most LLMs don’t read text as words or characters. They process tokens. What if we skip subwords entirely and model raw bytes?

Neither is "better" everywhere. For clean, high-resource text, subwords are extremely effective. For spelling-sensitive tasks, noisy inputs, rare languages, or text where tiny character changes matter, bytes can be surprisingly attractive.

This is the core trade-off: Subwords: • shorter sequences • strong lexical units • efficient • tokenizer-dependent Bytes: • universal representation • robust to rare/noisy text • exact surface detail • longer sequences • more work for the model

There’s another subtle trade-off. Subword models get useful chunks "for free": ["photo", "graphy"] A byte model starts with: p h o t o g r a p h y It must learn internally that these bytes form meaningful larger patterns.

Self-attention scales quadratically with sequence length. So a 5× longer sequence can mean 25× more pairwise position interactions in the attention step. That’s the trade-off: Bytes buy universality and robustness, but longer sequences are not free