kshv

Sic transit gloria mundi & ... I also work on long horizon agents

[La-Ndlovu @destiny_ndlovu]

There should be an app where you upload your bills and bored billionaires swipe right to pay them.

[Nikita Bier @nikitabier]

It’s time to turn off the internet.

[kshv @Fainterstar11]

Do you remember when you joined X? I do! #MyXAnniversary

[kshv @Fainterstar11]

@alexandr_wang India release when ? 🫤

[kshv @Fainterstar11]

@alexandr_wang India release when ? 🫤

[kshv @Fainterstar11]

Lol

-----BEGIN RSA PRIVATE KEY-----
4
-----END RSA PRIVATE KEY-----

[Haiyu Wu @HaiyuWu1]

We are trying to build a free LeetCode for AI researchers. You can learn how to code core ideas from important papers across various areas (e.g., LLMs, JEPA, RL, diffusion), contribute to the problem bank, and importantly, post your thoughts to find like-minded people for collaboration or job opportunities.

Of course, the current version is still far from mature, so anyone who wants to help us with this project is welcome!

[kshv @Fainterstar11]

The strawberry emoji 🍓 is 3 tokens with Opus 5.5, but only 1 token with Gemini 3.8 Flash.
​LLMs use different tokenizers, which can lead to meaningful differences in token count.
​Anthropic uses a proprietary Byte Pair Encoding (BPE) tokenizer with a smaller 65k vocabulary, whereas Gemini uses a SentencePiece tokenizer with a larger 256k vocabulary.
​There won't be much difference in token quantity for common English words (these should all exist in the vocab), but it could make a difference of 40%+ for certain use cases, such as with non-Latin characters (e.g. Devanagari script) or obscure code syntax.
​This is yet another reason cost per token isn't a reliable metric for understanding true costs. 'Cost per token' is misleading without knowing tokens per unit of work.

[sadkat @sadkatwt]

Interview panel for hiring a junior dev

[Ryan Florence @ryanflorence]

ngl "let there be light" sounds like a prompt

[kshv @Fainterstar11]

agent HCI problem nobody names: you don't know when to trust the agent until it's already done the thing.

[kshv @Fainterstar11]

@HosseinGorji021 B, C and D are one answer: the agent optimized 'tests pass', not 'it works'. A is the only one where the reward and the goal agree

[kshv @Fainterstar11]

@bravish_ hacks got stronger down the chain after the source left. does a verifier agent placed mid-chain stop it, or does the chain just launder it?

[kshv @Fainterstar11]

@sukin_s the 4B beating its 35B teacher is the wild part. is the edge from RL on execution reports, or just narrower scope (single-step revisions)?

[kshv @Fainterstar11]

@bravish_ fine-tune on one track, chance on the other. sounds like the detector learned the dataset, not the hack. any held-out-source split?

[kshv @Fainterstar11]

@Joshua_WD 3 held-out tasks, written by GPT-5.5 and judged by GPT-5.5. how much of the 70% is just the judges liking their own family's tasks?

[kshv @Fainterstar11]

@HosseinGorji021 agents is a big bucket. orchestrating several, or one agent with too many tools?

[Abhijeet Dipke @abhijeet_dipke]

I had requested to let me speak to my cousin who lives with my grandmother. Asked for my phone just for a couple of minutes so I could at least take out the number of my cousin.

Why Delhi Police didn’t let me use my phone? I had also asked for a call with lawyer which was again denied.

At 3:30 in the morning, Delhi Police asks me to call my father and inform him that tomorrow I’ll be going in jail for a long time. The same father who has just lost his mother. How much more trauma did you want to put my family through after we already lost my grandmother?

What was the purpose of trying to call my parents at f-ing 3:30 in the morning? You wanted to torment them?

[kshv @Fainterstar11]

@bravish_ does the 40k synthetic set come from real agent runs or generated specs? curious whether the exact-match gap is the data or the task itself