Nik Cubrilovic

I'm very smart. Ask me anything.

@leerob is it too soon to make a kimi 2.5 / composer 2 joke with this? :)

Today's agenda: training my daughter for the big leagues

@Mikebps @SinaHartung his dick stopped working before he turned 18

@SinaHartung You don’t think he’s getting laid at all? Be for real.

my favorite things about clavicular and the manosphere is that they care about the aesthetic of the possibility of having sex but don’t ever get laid

@chribjel their business model is revenue from licensing at the top-end

i wish the open weight models like kimi/glm/minimax would have a license so all derivative models also would have to be open weight

Fiverr is dramatically losing rank positions. SEMrsuh est. traffic is so wild per keyword - its pointless looking at. The loss in positions is key

@ns123abc @sama man you really post some shit. there's no model atm that touches 5.4-pro. your paymasters at X put out an 8th-tier model after almost a year of delays and couldn't even beat chinese labs with 10% of the market cap and funding

OAI is just incredibly lost. @sama realized that without the powerful model, his products are ass Altman + bad product == Loopt no one at OAI is capable of SOTA research, not anymore at least

@nitzukai after USB-C we'll have USB-C++ and then USB-Rust

is USB-C the final evolution of ports or is there any possible design that could replace it in the future before we go fully wireless

@_chenglou @theo and with 6x price for 2.5-3x speedup suggests speculative decoding

@theo The fact that the KV cache isn’t shared when turning opus fast mode should be a good hint prolly

Can someone smarter than me explain how these “turbo mode” things work on like an implementation level? Are they using better GPUs? More GPUs? Reserved allocation during generation?

@auchenberg Google need to acquire gstack - they won't even need to rename it

Anthropic got Bun. OpenAI got Astral. The big coding labs are the new acquirers of developer tools.

@HanchungLee @samuelcolvin good luck adding numbers

@samuelcolvin good luck training models using typescript!

PYTHON + RUST. Python inside the sandbox, Python (and some Rust) outside the sandbox. Typescript for frontend developers trying to stay relevant, like Mastra. By the way, the SF bubble is the one place where TS is popular for AI: `openai` package has 53m weekly downloads on PyPI, and 10m on NPM. AND THAT GAP IS WIDENING - used to be 4x, now 5x.

@jonathan_wilke @nextjs my CPU cores thank me every day for migrating to vite + tanstack start

Holy crap... with the @nextjs 16.2 update the next server is now using 10GB+ of memory and my MacBooks fan is running constantly (before it never even turned on). What's going on here...

She looks just like someone else but I can’t quite put my finger on it…

@NelsonXLee @vercel @anything @supabase I can see vercel eventually building their own network and hardware

Gonna flesh this out in a 2k-word article later this week. Companies like @Vercel, @Anything, and @Supabase are great $1B companies, but they’ll never be $100B companies. Their business model prevents any power law.

@nombreuten @tleilax___ so what you really want is anomaly detection across FIRMS data - pretty sure this is how google are doing their bushfire detection.

@tleilax___ Interesting until you realize “fire” is common everyday at oil-gas facilities from these imagery. Just check previous dates on NASA Firms… Middle East facilities always full of red squares 🫡🟥

Some building near the Riyadh refinery got hit. Might be the pipeline leading to it. Intense fire, detected 90 minutes ago.

@steveruizok in the 90s I used to take a 2 hour train ride to a library to take notes from programming books. I would have preferred what we have now.

It does bother me that it costs money to code now

@rachpradhan true bench would be fastapi builtin, fastapi pydantic, fastapi dhi, turboapi dhi - most of the gains would be squashed while maintaining backwards + forward compat

@rachpradhan most of the gains here would be dhi serialization over builtin python (which is sloooow): #L104-L194" target="_blank" rel="nofollow noopener">github.com/justrach/turbo… doesn't take advantage of pydantic / rust based serialization which landed in 0.130.0: #json-performance" target="_blank" rel="nofollow noopener">fastapi.tiangolo.com/advanced/custo…

I replaced FastAPI's entire HTTP core with Zig. Same decorator API. Same Pydantic models. 7× faster. 47,832 req/s vs FastAPI's 6,800. 2.09ms p50 latency. Introducing. TurboAPI. Here's the story..

@__trackertor__ @___4o____ it worked

@___4o____ Counting on mass to be stupid

These people have terrible opsec lmao

@majamediaco this is down the road from me and it becoming an IG spot has been a horror

this isn’t actually even Sydney, but it does look like this

@wesbos perplexity to acquire left-pad

Think Deno is being acquired by OpenAI? ~8 of their employees posted about no longer working there today, one said two weeks ago they resigned, none of them said layoffs. All on bluesky, while Ry is posting about ChatGPT water usage not being that much Always been a huge Deno supporter, curious to see so many leaving

TNT Sport just interrupted Sporting CP Vs Bodo/Glimt to say Netanyahu has died. They announced it like it was a goal in another game.