Bitduke

quite an important addition to the post about building a custom inference engine for deepseek v4.1-flash (from scratch) turns out the prompt makes a difference, since those numbers came from one specific prompt where the engine averaged 5.6 accepted tokens per step, so you

interesting, even MEV bots are now basically a 5-minute job with Fable 5.2, stuff that used to to take an expert days, competition is only getting tougher

wild

yeah, another reminder that you don’t need the most powerful model for most tasks

if you haven’t checked out Claude Duet yet, now’s a good time. everyone on the US waitlist got access

Cursor reported that a read-only key leaked by Fireworks let an attacker access old coding agent logs, find more credentials and eventually plant a malicious copy of the Claude Code extension in Cursor’s marketplace cache pretty unsettling that reading logs could unlock that

interesting, some researchers are going beyond tweaking inference settings and using ai to build their own engines around specific models and GPUs here, they report running Deepseek 15% faster than SGlang on their setup by building a small engine where they could test

it’s funny that llms have completely changed in the last 24 months, and people still can’t wrap their heads around the idea that progress might speed up. perhaps things really are moving too fast for a human to comprehend Dwarkesh, who’s usually careful not to overestimate

so it seems Claude Mythos wasn’t a one-time deviation, just a point on a superexponential curve

Dwarkesh was the guy who kept pushing back on short ai timelines, and now his own model has the RSI loop starting in 2027, so the gap between the skeptics and the RSI-2027 crowd has shrunk to a year or two

lifted

it seems like every big lab is launching its own decision model, and people are actually using them decision models went from zero Vercel ai gateway requests in Aug to 26% in October (just 0.53% of total spend tho). almost all that volume went to typesafe ai’s Jev, but it helps

Claude Code now has custom themes

yeah, pretty much

so, it could’ve been ai-related after all (not a cyberattack tho) the FAA was rolling out SMART, its AI air traffic tool, when the nationwide ground stop hit (SMART is basically the FAA’s ai tool that helps plan flight routes and reduce delays (ironic if true))

Elon Musk announces that he is signing a petition to bring Jensen back to Nvidia

a communication issue across U.S. air traffic control centers has prompted the FAA to halt departures nationwide the last comparable nationwide ground stop was on Jan 11, 2023, when its flight safety notification system failed planes already in the air can still land, while

more open-weight models are always good (especially if you can try them for free)

Google’s cheaper Gemini is pretty good at selling cans of coke it turned $500 into roughly $14.5k over a simulated year and took #3 on Andon Labs’ Vending-Bench 2. that means finding suppliers, negotiating prices and keeping stock moving over thousands of interactions, which