Brian Fioca
2.8K posts
Brian Fioca
@bfioca
Just a person, mostly. Applied Evals @ OpenAI, GPT-whisperer, co-founder of RescueTime, YC ‘08 alum, also likes 🎺 and 📸. Ideas very much my own.
Seattle Katılım Şubat 2007
514 Takip Edilen2K Takipçiler
as a kid i played sports video games exclusively on the hardest difficulty & i’d get destroyed for weeks. then one day i wouldn’t. i love starting where it hurts & then repeating until it doesn’t. i prolly treat every problem or goal in my life this way.
i even read almost everything backwards too, e.g. i usually start by reading the end of a book then work my way backwards (same with almost all articles).
English
@bfioca @johnawahba this kind of ontological elegance is something that is really hard to come up with a good eval for
the closest mechanical thing is probably just lines of code tbh
would love to use a model that was RL'd hard for code golfing
English
had managed to get myself down to less than 10% of my codex weekly rations with 3 days remaining
had to make do with gpt-5.4-medium for the last few days to subsist
but today we're back baby
English
5.4 xhigh is overkill for a lot of things and might actually overcook your task, unless you're going as hard as this
Grant Slatton@GrantSlatton
@bfioca @johnawahba today's work was coming up with this batch optimization system that has a whole associated ontology the minimum representation is probably like 12 classes that interlock in a complex way 5.4-high just struggled and failed to find the right factoring x.com/GrantSlatton/s…
English
@bfioca @johnawahba today's work was coming up with this batch optimization system that has a whole associated ontology
the minimum representation is probably like 12 classes that interlock in a complex way
5.4-high just struggled and failed to find the right factoring
x.com/GrantSlatton/s…
Grant Slatton@GrantSlatton
graphical analogy: current LLMs have bad taste at where to draw abstraction boundaries suppose data flows through your system like pic 1 LLMs tend to make kind of arbitrary abstraction boundaries like pic 2 whereas a human might do more like pic 3
English
@GrantSlatton @johnawahba super interesting! I'll pay more attention to this in my evals
English
@bfioca @johnawahba i tried high all today and i'm switching back to xhigh… high is just not thoughtful enough for my work and i found myself just wasting time trying to get it to understand and frustrated
English
@GrantSlatton @johnawahba High usually gets me there for most things. Xhigh for when I walk away for more than a few hours
English
VR for me continues to be too alienatingly isolating as someone who doesn’t live alone.
Kevin Roose@kevinroose
“The Metaverse” was a 5-year experiment conducted by Harvard University for the purpose of transferring wealth from Meta shareholders to employees who pretended it was fun to go to meetings in a VR headset. We thank you for your participation.
English
Super good call - I hated useEffect even before AI 😅
Eno Reyes@EnoReyes
Alvin has a great writeup on why our team decided to ban the use of useEffect from our codebase. Sometimes you have to make semi-controversial calls to ensure Agents can operate with maximum effectiveness in the codebase. This is one of those calls!
English
Brian Fioca retweetledi
we shipped a new version of 5.3 instant to chatgpt yesterday. 5.3 was unintentionally pretty annoyingly clickbait-y. it's better in yesterday's model and we're going to keep stamping that behavior out. keep the feedback coming!
help.openai.com/en/articles/68…
English
Don't sleep on 5.4 Mini - in most of my tests it scores better than original GPT-5 at launch.
Factory@FactoryAI
GPT-5.4 Mini has an impressive cost-to-capability ratio and punches above its weight on practical multi-step engineering work. Other strengths include: • Strong step-by-step planning • Cleaner, more reliable tool use • Better persistence through debugging and blockers • Clear gains from higher reasoning effort
English
Brian Fioca retweetledi
Subagents are now available in Codex.
You can accelerate your workflow by spinning up specialized agents to:
• Keep your main context window clean
• Tackle different parts of a task in parallel
• Steer individual agents as work unfolds
English
Brian Fioca retweetledi
gpt-5.4 has ramped faster than any other model we've launched in the API: within a week of launch, 5T tokens per day, handling more volume than our entire API one year ago, and reaching an annualized run rate of $1B in net-new revenue.
it's a good model, try it out!
English
@viemccoy @way_opener People can adapt pretty fast. For me the singularity is when things get weirder faster than we can adjust.
English
@way_opener its going to get so weird that we're going to have to take a step back and talk about how weird things have gotten
English
I don’t really believe in “the singularity” per se but there is something of an event horizon, terence mckenna did kinda call it wrt rapidly increasing novelty. Things will get weirder at an exponentially increasing rate but our perception of time will somewhat slow to compensate
English
I’ve spent more than a decade meeting startup founders in Seattle, learning about their businesses and sharing their stories. Now I’m ready to work alongside founders and help boost the next generation of technology startups at the AI2 Incubator as the new director of AI House.
English
I’m excited to share that I'm joining @OpenAI as AI Deployment Engineer for Startups.
It has become clear that software is changing fundamentally. More and more, founders can shape the tools they use directly with AI, and the ceiling of what can be built keeps rising faster than ever.
What makes OpenAI special is the scale of its ambition: to push the frontier of intelligence while making it safe, accessible, and actually useful for the builders who will define the next era.
After a year as founding AI engineer at startups, I understand the startup space from a deeper perspective. I hope to leverage this knowledge and take it to the next level at OpenAI - helping founders bring advanced models into production safely and at scale.
Grateful for this moment. Open to connecting with fellow founders and builders!🚀
English