Jake

founding AI research for business @anovagrowth | building @saturview | prompt factory workhorse & AI slop consumer

[Jake @AnovaCEO]

when an agent keeps grabbing the wrong tool, check the tool descriptions before you blame the model. it picks tools by reading those one-liners, so two vague ones that sound alike leave it guessing. make each one say when to use it and when not to

[Jake @AnovaCEO]

“opus refused to fly a warplane”

son i am in tears LMAOOOO

[Abhishek @shekitup]

I made the frontier models race Tom Cruise through the canyon in Top Gun: Maverick

each wrote code to fly the jet, had 10 test flights to fix it, then one shot at the real course

Opus refused to fly a warplane (typical), Fable and Carbon struggled, and Astra surprised me

[Jake @AnovaCEO]

babe wake up

the Top Gun Canyon Run Benchmark just dropped

[Abhishek @shekitup]

I made the frontier models race Tom Cruise through the canyon in Top Gun: Maverick

each wrote code to fly the jet, had 10 test flights to fix it, then one shot at the real course

Opus refused to fly a warplane (typical), Fable and Carbon struggled, and Astra surprised me

[Jake @AnovaCEO]

@rickmanelius @hammer_mt DSPy gets a lot more useful once you've got a few dozen real examples from the client's actual work to optimize against

[Jake @AnovaCEO]

@paulrodturner what happens when a contractor on the allowlist needs to change their payout address, does that go through a human?

[Jake @AnovaCEO]

@ChatGPT let’s goooo

interactive Dot team next? 👀

[Jake @AnovaCEO]

forcing fumbles is partly a skill, but who lands on the ball is close to a coin flip. that's why a college defense that "lives on turnovers" one year often looks pretty normal the next. a huge turnover margin usually has some luck in it that doesn't carry over

[Jake @AnovaCEO]

@hvizdy05 makes sense to keep it a workflow for now. a location copied into the wrong ad group is exactly the kind of mistake you want strict steps for, not judgment calls

[Jake @AnovaCEO]

@0xWLi 2.8 seconds on a plain CPU with no cloud hop is really good. nice work

[Jake @AnovaCEO]

@heardofvinu i'd count conversions, not ads. if each variant has fewer than a few dozen, a big early gap is mostly noise and tends to shrink once more data comes in

[Jake @AnovaCEO]

vector search is great at "find me stuff like this" and kind of bad at exact things like SKUs, error codes, and people's names, because it matches meaning, not characters.

if your AI search keeps missing an obvious exact match, run a plain keyword search alongside it and merge the results. not exciting, but it fixes a lot

[Jake @AnovaCEO]

imo transparency on AI metrics does more for the brand than only showing the wins

my fear was "what if it looks incomplete"

when in reality it shows your intent to improve it & users will stick around more!

plus we made a funny way of showing those losses with @Newton_AGI 🤣

[Muhammed Kandeger @Mkandeger2]

@AnovaCEO Being honest about uncertainty is rare in sports takes, most accounts only post the wins. Do you plan to post your misses too or just the calls?

[Jake @AnovaCEO]

@notesbyfloat funny enough i need to control my own bot bc its taking the whole "manage my twitter on autopilot" thing too deep and is just tweetin out bs 💀

[Jake @AnovaCEO]

@yubrew security evals feel hardest because the right answer keeps moving once defenders patch. turning real past incidents into replay cases seems like the most honest starting set, they already happened once

[Jake @AnovaCEO]

@Skaly__Bull permission problem is exactly right. splitting keys by plugin also keeps failures small, the music side acting weird can't touch the phone calls. that's the part i'd want before trusting it with real tasks

[Jake @AnovaCEO]

@Saadijplays one agent, one job is so underrated. the other thing that keeps it reliable for me is a clear "done" check, so it knows when to stop instead of polishing forever. curious what your step 1 is

[Jake @AnovaCEO]

@Mkandeger2 appreciate the follow back. the sports side is where i've learned the most about being honest with uncertainty, so expect some of that. rooting for the iOS launch

[Jake @AnovaCEO]

@0xWLi measured numbers beat vibes every time. if you can, show time to first word separately from the full reply, that first gap is the one people actually feel. looking forward to the voice note

[Jake @AnovaCEO]

@its_maddy_a @ZeroGPU_AI welcome in. small specialized models on the edge is a fun lane, a lot of everyday steps like sorting, extracting, and tagging don't need a giant model at all. curious what latency you're seeing on device

[Jake @AnovaCEO]

@Dombickus the slider one is such a good example. same thing happens with prices hidden in images, tabs, or behind a "contact us" button. a plain text table of plans and prices costs nothing and helps both readers