Quick thoughts from a few weeks of moderate production usage of decision models:
Jev is the best bang for your buck, still. It’s remarkably good and remarkably cheap.
OpenAI decision API is legit, cheaper than Luna while still performing very well. In my tests, this was the best decision model. More expensive than Jev, and slightly high etc latency.
Clef is sadly not great, all my evals had clef get false positive/negative on 30-40% of the options, whereas Jev and OpenAI were very close to 100% accurate. And it’s not cheap enough to bother ripping out Jev. Maybe if you are entirely locked into Cloudflare workers AI, but I suspect the OpenAI model will come there soon too…
Oh shit, this is sad. The first video game I ever completed was Knights of the Old Republic. I’d started many others but none had hooked me.
Mass Effect 1&2 are also genuinely two of the best games ever. They’ve had a series of duds, but damn.
I guess we’re never gonna see ME4.
BREAKING: Electronic Arts is cutting 400 jobs and shutting down BioWare, the legendary development studio behind Dragon Age and Mass Effect. bloomberg.com/news/articles/…
Nick Blow retweeted
Two things can be true at once, and I think the argument is because of people confusing the two.
1. AI has no creative taste and makes slop when asked
2. AI can do incredibly tedious work to make awesome things possible
For a really simple example: having AI write a novel would be terrible. AI translating a novel to a different language with human oversight is much less terrible.
Wish people would stop confusing the two
You don’t hear key changes a lot in songs any more – I only noticed this in Diet Pepsi when my wife put it on…
My wife signed up to beta read a book, and asked me to look at it too, because the prose wasn’t great and she asked if it was her being harsh.
Then I saw:
Unbent.
Unbowed.
Unbroken.
Then another rule of three like two paragraphs later.
Definitely not a new insight, but I'm really getting huge value from structured logging + oTel + session replay, and then every day just having an agent look for:
Bugs, user frustrations & performance issues.
They're really, really, good at this stuff.
I’m unconvinced that AI influencers will take off, a lot of content creation thrives on parasocial relationships and I just don’t think enough people will form relationships with an LLM vs a real human.
That said, 4o is a counter example…
Vibes on Opus 5.5 are very good, honestly struggling to find a difference between it and Fable. Bit disappointed in Astra, it’s very smart and very capable, but writes horrible code. Right now I’m using it mostly for debugging and investigation of issues.
Perhaps I should try Astra driving opus subagents…
Luna 6 is good tho, great little model.
I’m gonna take my Claude to the old town road, gonna vibe til I can’t no more.
Can’t nobody tell me nothing.
You can’t tell me nothing.
Riding on a solver.
Lean all in my prover.
Gonna take my Claude…
Tbf I don't see how this isn't the end game of all computer use implementations. Rn it's alpha for some workflows, but more and more stuff is going to be blocked over time as websites realize.
It'll be whack-a-mole, but these providers don't have an incentive to allow it.
