Fanfulla

Applied AI Engineer | @cognition and @n8n_io Ambassador | Growth & Product | Data-Driven CX

[Fanfulla @Fanfulladev]

This is Fable 5.2 xHigh on Devin

52 minutes to make this video of a monkey riding a red Vespa in Rome near the colosseum

Make no sense how good it is, what did you run yet on it?

[Fanfulla @Fanfulladev]

@Tesla_Optimus_K Manual approval for every dev is fine with me for now, nobody wants a random script sending 50 pound carts around a neighborhood

[Fanfulla @Fanfulladev]

@vllm_project @SemiAnalysis_ @NVIDIAAI @MiniMax_AI Counting cached tokens feels fair here, agent loops reread the same context all day long!

[Fanfulla @Fanfulladev]

@a_lungeanu Hooks from a stranger's repo, what could possibly go wrong

[Fanfulla @Fanfulladev]

@meetup_milo Customer ID 12345 strikes again!

[Fanfulla @Fanfulladev]

@Shadowfetch A fake AWS where agents can break stuff all day and no bill shows up, love this

[Fanfulla @Fanfulladev]

@KostyaAI Scariest rework is the one nobody spots, an agent that quietly changed what the app does still counts as accepted

[Fanfulla @Fanfulladev]

@llmxlabs That works out around 28 tok/s per request at 108 clients, fine for batch jobs but long thinking traces will feel slow for whoever is chatting

[Fanfulla @Fanfulladev]

@LesOttolenghi Member rates hidden behind a login popup are going to hurt a lot here

[Fanfulla @Fanfulladev]

@EverydayEddie_ One tab gets my vote, splitting Chat and Work always felt pointless

[Fanfulla @Fanfulladev]

@RoundtableSpace Auto finding Ollama is my favorite bit of this release, local models should need zero setup

[Fanfulla @Fanfulladev]

@solcardai 104B active per token is still heavy, even the fast hosts will have to work hard to serve this one

[Fanfulla @Fanfulladev]

@doronkatz Cost per task belongs next to every agent demo, nobody ever shows how many tries the deck took

[Fanfulla @Fanfulladev]

@metamitya is one of the three a test suite the agent has to pass before anything deploys?

[Fanfulla @Fanfulladev]

@stroopstack @thsottiaux Do your sessions actually get past 260k before compaction kicks in?

[Fanfulla @Fanfulladev]

@MattKimball_MIS @AMDServer Agents burn so much time in tool calls and sandboxes, good to see CPUs getting love again

[Fanfulla @Fanfulladev]

@The_havix Keeping a list of what's still unverified is smart, agents love treating their own guesses as facts

[Fanfulla @Fanfulladev]

@Ishant23April Scariest case is an agent with shell access, the trigger could just sit in some file it reads

[Fanfulla @Fanfulladev]

@leocardz Monday morning humans would never pass these evals

[Fanfulla @Fanfulladev]

@Moretti_ale90 I'd want to see how many free brains turn into paid ones every Monday

[Fanfulla @Fanfulladev]

@danblake Nice test for this: give an agent the lowest role and let it poke the API to see what it can still reach