qasim

here's how to bypass cloudflare turnstile/vercel botid/etc gated endpoints. not a lot of people know about this, and i decided i would make a quick tutorial for it. motivation: @thorchh add vercel botid to gate the vote endpoint on hackelo.com. i was really into webscaping & cyber security in middle school, and this brought back a lot of memories. hope people learn something from this!

who else can only focus when the right song comes on?

why aren't people running ads on their resume yet?

learning with ai is placebo

well this is very nice.

liking coding is newgen now.

never forget: your maximum is someone else's minimum.

ig they saw my tweet 😋

here's my secret playbook to effortlessly stay productive at 4am 👇 - sing along to taylor swift & disney

the next robotics frontier is removing stains from clothes. why? because this task 1. has generally very low success rate (even for humans) 2. requires incredible precision in movement and also visual inspection, which generalizes to many different use cases. 3. requires good analysis and thought process, showcasing robot autonomy. the list goes on... i genuinely believe any robotics lab able to solve this will be a unicorn.

no one in sf knows how to dap properly

i miss poutine 😔

codex 😭😭😭😭

i feel like the problem isnt as simple as routing to the 'best model'. i'm not sure how the recent model fusion models work but, there's so much to account for. for example, if we blindly switch models, we lose the benefits of prompt caching, causing our model routing to be MUCH more expensive at a not equivalent intelligence/speed increase? i could only think of a master-slave relationship (shoutout systems design), where the communication threads are not the same, but more the main model routes a task as a subagent to not break the cache but to get that extra gain... in that case, would it still be considered a 'router' or 'fusion'? or more just harness engineering? i haven't looked too much into it, but perhaps im missing something?

i am genuinely amazed, wtf

i can't begin to imagine what llms will look like next year.

how can a small startup (like us) afford to run on tbench v3 😭 Opus 5 (max) — 43.5% ± 1.7% — $6.0k GPT-5.6 Sol (max) — 34.6% ± 1.6% — $4.0k Fable 5 (max) — 34.1% ± 1.7% — $7.1k Opus 4.8 (max) — 21.1% ± 1.6% — $5.7k GPT-5.6 Terra (max) — 20.8% ± 1.4% — $2.5k Grok 4.5 (xhigh) — 15.7% ± 1.5% — $960 Sonnet 5 (max) — 14.6% ± 1.5% — $7.9k GPT-5.6 Luna (max) — 14.3% ± 1.3% — $1.6k GLM 5.2 (max) — 4.6% ± 1.0% — $3.8k any labs wanna give us some 💳😋

conspiracy theory: opus 5 is so bad, so they prevent benchmarking on it by rerouting u to opus 4.8.

sama rlly looksmaxxed their new models, 5.6 sol has been really good with ui.

terminal bench 2.1 is exciting! previously codex was #1, but we recently found out that it was actually 4x more expensive than fable. so it's actually CHEAPER and BETTER to use fable than it is to use codex gpt-5.5. these new parameters are more exciting, scores aren't the most important anymore, its should be a index based on hacks, score, and cost.

i have 16 chances to make a billion dollar company.