@GergelyOrosz I do not entirely get the ‘loop” part. I usually prompt the model to work until the flaky tests are deterministic again or ran out of ideas - and document everything on the ticket as comments. Is this considered a (do-while) loop?
@sean_a_mcclure Nobody mentioned dictionaries. I am willing to die on the hill that if a definition needs a context that is open to interpretation then it is, by definition (you might want to look up the etymology of the word), not a definition.
@AdamRackis Wait until a designer teaches you to tell apart good and bad kerning. (Do not google it if you have the good fortune of not knowing what it is).
Your entire life will change when you stop assuming people know how you feel about them. Tell your friends you’re proud of them. Text your parents you love them. Compliment your coworkers. Say the kind thing while you still have the chance. It’s something you’ll never regret.
@unclebobmartin Space = space (pun intended) + electricity is virtually free and unlimited, launch capacity is three and improving fast. Gigafab will solve the chip&memory issue. What am I missing here?
@mbateman Both have it’s benefits. I believe if you completely miss one (never playing with the kids on the beach, or never let them play freely while you just observe ) then their childhood is less rich than you could’ve made it.
@GergelyOrosz I had a CTO who wanted to tie bonuses to LOC. Withdrew the idea when I shared that the challenge is to have the same functionality with fewer lines and I can make an arbitrary large Hello World. No reason you can’t do the same with PRs.
@valigo I keep the bozos out by asking for a 100 line real code they wrote. Most just ghost the process without sending anything. Whoever send a good 100-liner, have a 80% chance to ace the interview. One sent an open source repo he contributed to. Got an offer at 1st the interview.
@quxiaoyin Same here, altough the 5h token budget helps on not to stay locked in for too long. Claude codes, Grok manages the requirements, I babysit both. 1 super intense hour 2 or 3 times a day on top of the regular work is a lot.
@davepl1968 There is an expectation that you do your part on checking if your solution is not infringe any patents. You are not obliged to, but need to understand what you go into. (Managed product dev and innovation for 6 years, this is the least fun part.)
@GergelyOrosz@EngGuidebook Just ordered on https://t.co/ele8dCHqwV. I anticipate the contemt will be gold and also can’t wait to see the translations you collected/made… everyone I know calls those terms the same around me, just with Hunglish promounciation. :-)
@KettlebellDan Asked Grok on Monday how to choose a good bike for me. Went down the hole as far as to ask it to review the marketplaces and make a shortlist. Bought the exact one it suggested as #1 choice, today I rode it for 20 miles. This is how deep it gets eventually. You have been warned.
@TheTeslaSignal This is not how the staged rollouts work. In reality you recalculate the cadence, next batch size and schedule based on the feddback and info connected.
SpaceX has almost finished writing V1.0 of an in-house AI training stack in C that exact-maps to 220k GB300s with 800G NICs, making heavy use of pipeline parallelism and getting as close to bare metal as possible.
The potential speed improvement vs JAX for large training runs is over an order of magnitude.
@KettlebellDan The speed is impressive and it will catch up or take the lead. Now it can fall in love with it’s own explanation why the code is ok but the environment is not (asking me to fix) while Opus 4.7 1M figures out the issue and fixes it. Will try again when the big model lands.