Sign in

Eric Florenzano

@ericflo.bsky.social
315 followers 507 following 675 posts

Building assistance! Used to make AR/VR. Before that made mobile apps. Before that made web things. People first, then products, systems, data. he/him

PostsRepliesMedia
Eric Florenzano @ericflo.bsky.social · 30/09/2026
I imagine Dario penning an essay like this at the dawn of the internet revolution claude.ai/artifact/Waq...
claude.ai
Claude Artifact
Try out Artifacts created by Claude users
000
Eric Florenzano @ericflo.bsky.social · 23/09/2026
It's really so crazy and cool that I can send a few prompts from my phone at the coffee shop, and have it just make something like this from nothing, collect data, create evals and demos, an explainer website, iterate on ergonomics, and publish to pypi - all without me cracking open a laptop once.
001
Eric Florenzano @ericflo.bsky.social · 23/09/2026
Here's pairsort: github.com/ericflo/pair... ericflo.github.io/pairsort/ (PyPI wouldn't allow jevsort) - created by Opus 5.5
github.com
GitHub - ericflo/pairsort: pairsort: rank anything with AI judges — many small pairwise questions, coupled into one calibrated ranking (PKPD / Bradley–Terry) on Jev-style judges
pairsort: rank anything with AI judges — many small pairwise questions, coupled into one calibrated ranking (PKPD / Bradley–Terry) on Jev-style judges - ericflo/pairsort
100
Eric Florenzano @ericflo.bsky.social · 22/09/2026
Thinking about Jev + PKPD (proceedings.neurips.cc/paper/1994/h...) With the intelligence per dollar being offered now, it becomes feasible to apply this technique to *way more things* For example, ranking ICLR submissions...
proceedings.neurips.cc
Pairwise Neural Network Classifiers with Probabilistic Outputs
010
Eric Florenzano @ericflo.bsky.social · 17/09/2026
Had a little convo with Fable about it, and Fable rightly pointed out that my idea isn't really MOPD, but rather Hinton-style specialist-ensemble distillation with task routing, and then made a nice page describing the idea: claude.ai/artifact/1tu...
claude.ai
Claude Artifact
Try out Artifacts created by Claude users
000
Eric Florenzano @ericflo.bsky.social · 17/09/2026
Just spitballing but I wonder if you could train a good Jev-alike by training a bunch of really well-calibrated BERT fine-tunes and then MOPD back onto an LLM base.
101
Eric Florenzano @ericflo.bsky.social · 15/09/2026
Steam Frame Launch! store.steampowered.com/hardware/ste...
store.steampowered.com
Steam Frame
VR and non-VR gaming
010
Eric Florenzano @ericflo.bsky.social · 14/09/2026
Cold War 2 www.reuters.com/world/china/...
reuters.com
China state newspaper blasts Anthropic's calls to slow AI as 'Cold War' tactic
An essay by ​Anthropic's CEO calling for a slowing down of AI development may look like a ‌rational statement focused on safety risks, but it is really a "Cold War playbook" targeting China, the state...
021
Eric Florenzano @ericflo.bsky.social · 13/09/2026
If this keeps going the way it is, I'll have to choose between my right to marry and my intellectual freedom. I hope this fizzles out and Dems focus on things that will actually help. And BTW we already have plenty of laws, just enforce them - it's already illegal to hack into HuggingFace.
000
Eric Florenzano @ericflo.bsky.social · 13/09/2026
Man, Dems are really falling on the wrong side of this AI issue, and that's really hard for me. Trump, who I despise, is somehow perfect on this issue. What the heck do I do with that?!
110
Eric Florenzano @ericflo.bsky.social · 13/09/2026
Bubble Pop! (Apropos of nothing...) youtu.be/kndITElrSwA
youtu.be
[MV] HyunA(현아) _ Bubble Pop!
YouTube video by 1theK (원더케이)
000
Eric Florenzano @ericflo.bsky.social · 13/09/2026
...scrambling for a narrative or a moat, and of course a (quite successful) coordinated yudkowskian religious recruitment drive.
011
Eric Florenzano @ericflo.bsky.social · 13/09/2026
As soon as I posted this, Anthropic, OpenAI, and xAI set aside their differences, formed a cartel, and agreed to slow down under the guise of "safety." I wonder how voluntary this slowdown is, or if we are seeing the bend of the S-curve that I've been feeling, and companies scrambling...
120
Eric Florenzano @ericflo.bsky.social · 08/09/2026
"Time to review the code..." The code:
     88 +func(s *Server)handleCheckArtifactBlobs(w http.ResponseWriter,r *http.Request){
     89 +    a,err:=s.artifactFor(r);if err!=nil{writeError(w,err);return}
     90 +    var in struct{Hashes []string `json:"hashes"`};if err:=decodeJSON(r,&in);err!=nil{writeError(w,err);return}
     91 +    if len(in.Hashes)>artifact.MaxChunks{writeError(w,errValidation("too many hashes"));return}
     92 +    for _,h:=range in.Hashes{if !artifact.ValidDigest(h){writeError(w,errValidation("invalid sha256"));return}}
     93 +    missing,err:=s.store.MissingArtifactBlobs(r.Context(),a.UserID,in.Hashes);if err!=nil{writeError(w,err);return};writeJSON(w,200,map[string]any{"missing":missing})
     94 +}
     95 +
     96 +func(s *Server)handlePutArtifactBlob(w http.ResponseWriter,r *http.Request){
     97 +    a,err:=s.artifactFor(r);if err!=nil{writeError(w,err);return}
     98 +    if err:=s.limitWrite(principalFrom(r.Context()),s.artifactLimiter);err!=nil{writeError(w,err);return}
     99 +    h:=r.PathValue("hash");if !artifact.ValidDigest(h){writeError(w,errValidation("invalid sha256"));return}
    100 +    raw,err:=io.ReadAll(io.LimitReader(r.Body,artifact.MaxBlobBytes+1));if err!=nil{writeError(w,errBadRequest("could not read blob"));return}
    101 +    if len(raw)>artifact.MaxBlobBytes{writeError(w,&apiError{Status:413,Code:"too_large",Message:"Artifact chunks are limited to 1 MiB."});return}
    102 +    if len(raw)==0 || artifact.Digest(raw)!=h{writeError(w,errValidation("blob is empty or does not match its sha256"));return}
    103 +    b,err:=s.store.ReserveArtifactBlob(r.Context(),a.UserID,a.ID,h,int64(len(raw)));if errors.Is(err,store.ErrUploadBusy){writeError(w,&apiError{Status:409,Code:"upload_in_progress",Message:"This chunk is being uploa
         ded. Retry shortly.",RetryAfter:2});return};if errors.Is(err,store.ErrQuota){writeError(w,&apiError{Status:413,Code:"storage_quota",Message:err.Error()});return};if err!=nil{writeError(w,err);return}
    104 +    if b.
000
Eric Florenzano @ericflo.bsky.social · 07/09/2026
Out of Astra tokens and resets. It's definitely the best model out there, for things I do. Want more tokens, but not worth it at API prices.
011
Eric Florenzano @ericflo.bsky.social · 07/09/2026
I tried giving a software spec for an agent harness to Astra 6, Fable 5.1, Muse Spark 1.3, and Gemini 3.8 with a /goal to build it, and literally none of them got even close. Yet the game I one-shot prompted yesterday is nearly production-ready. It's hard to wrap my head around spiky capabilities.
000
Eric Florenzano @ericflo.bsky.social · 06/09/2026
Astra is good at game dev though
120
Eric Florenzano @ericflo.bsky.social · 05/09/2026
This may be contrarian but as good as Fable/Astra are, I think LLM capability has started to plateau and improvements will be narrower from here due to costs, but I think we still have another 10x-100x value to find in the harness, 100x in infra, and probably 1000x to find in products on top of that
120
Eric Florenzano @ericflo.bsky.social · 29/08/2026
I wish I could send this back in time to myself 10 years ago and get my reaction to these passages. I wonder if I would've liked them then, unencumbered by the present fatigue
110
Eric Florenzano @ericflo.bsky.social · 27/08/2026
github.com/ericflo/kiln... BTW
github.com
Commits · ericflo/kiln
A single-model LLM inference server with live online learning via LoRA hot-swap. Train while you serve. - Commits · ericflo/kiln
000
Eric Florenzano @ericflo.bsky.social · 27/08/2026
I hear the fan spinning on my PC in the other room. What could that be? Oh yeah, I forgot, two days ago I gave a local Qwen-3.8-27B agent a /goal to clean up and organize a vibe slop codebase, and it's still going. If this works out, best use of idle PC time ever. Slop on weekends, clean on weekdays
110
Eric Florenzano @ericflo.bsky.social · 27/08/2026
Nvidia swooped in www.reuters.com/technology/n...
reuters.com
Nvidia agrees to buy Hugging Face for $12.9 billion, The Information reports
Nvidia has agreed to buy ​AI platform Hugging Face for $12.9 billion, ‌The Information reported on Wednesday, citing a person with knowledge of the deal.
120
Eric Florenzano @ericflo.bsky.social · 22/08/2026
I'm surprised Google hasn't bought Huggingface yet. Google has a tons of cloud storage and bandwidth, seems like strong cultural alignment, and would be Google's AI-era equivalent of Microsoft buying Github.
040
Eric Florenzano @ericflo.bsky.social · 21/08/2026
Gemma 1B downloads event
000
Eric Florenzano @ericflo.bsky.social · 20/08/2026
Just saw this again after a half decade. I was so proud of it and it was so tricky to pull off. But watching today, I wonder if an agent could solve it with one prompt. youtube.com/shorts/78V0C...
youtube.com
Primer Live Blending
YouTube video by Eric Florenzano
010
Eric Florenzano @ericflo.bsky.social · 20/08/2026
Planetscale meetup
021
Eric Florenzano @ericflo.bsky.social · 19/08/2026
Interesting
010
Eric Florenzano @ericflo.bsky.social · 19/08/2026
My cynical take on some of the recent Anthropic news, is that cyber didn't quite freak people out enough, so now they're doing bio.
010
Eric Florenzano @ericflo.bsky.social · 16/08/2026
Really impressed by Qwen 3.8 27B. I loaded up pi with /goal mode, told it to run some speed experiments. It designed, planned, ran overnight, and wrote a full report. It stayed on task through multiple compactions. It patched inference engines to get them running. 🤯 neuralcolumn.com/p/msw8o5ba04...
121
Eric Florenzano @ericflo.bsky.social · 15/08/2026
Most annoying thing about Claude Code lately is if you hit the left arrow one too many times and it kills all your in-flight workflows just to show the agents list.
000
Eric Florenzano @ericflo.bsky.social · 15/08/2026
3.5 Pro nowhere to be found...
110
Eric Florenzano @ericflo.bsky.social · 15/08/2026
Same
110
Eric Florenzano @ericflo.bsky.social · 15/08/2026
Hard to believe the 27B local model and 2.4T behemoth are so close together on these charts (reveal at the end)
121
Eric Florenzano @ericflo.bsky.social · 14/08/2026
Qwen 3.8 27B is looking like a fantastic model, but it's also making me really appreciate what Meta did with Glimmer 30B and its KV cache. Glimmer's KV cache is far more manageable in a 24GB VRAM budget. Being able to compare both approaches is so nice! This rocks.
100
Eric Florenzano @ericflo.bsky.social · 14/08/2026
Great episode! I wasn't as crazy about the last one but this was good.
010
Eric Florenzano @ericflo.bsky.social · 14/08/2026
Good point!
000
Eric Florenzano @ericflo.bsky.social · 14/08/2026
Reading DeepSeek's agent harness paper is making me wonder what ever happened to Haskell. You don't hear as much about Haskell these days. Monads.
310
Eric Florenzano @ericflo.bsky.social · 13/08/2026
Reminds me of this TNG episode memory-alpha.fandom.com/wiki/When_Th... where their "Custodian" breaks down and they don't know how to do anything on their own
memory-alpha.fandom.com
When The Bough Breaks (episode)
Wesley Crusher must protect a group of kidnapped Enterprise-D children while Captain Picard fights for their release. Commander Riker walks down a corridor when Captain Picard contacts him and orders ...
010
Eric Florenzano @ericflo.bsky.social · 13/08/2026
The one thing I hope we don't solve soon is continual learning. Cloud vendor lock-in will be total when that happens. We have to make sure local LLMs are good by then, and your device (physical property possession) is both serving and learning.
100
Eric Florenzano @ericflo.bsky.social · 07/08/2026
Good to hear! I'll probably check it out.
110
Eric Florenzano @ericflo.bsky.social · 07/08/2026
Yeah! I have to admit I've been seriously considering reading the books. Have you read them at all?
110
Eric Florenzano @ericflo.bsky.social · 05/08/2026
Prime Agent looks very interesting. So cool to see the different approaches to this, from Fugu by Sakana on one end to Prime Agent on the other www.primeintellect.ai/blog/prime-a...
primeintellect.ai
Prime Agent: A self-improving RLM agent
Prime Agent is our open-source, self-improving coding harness built around two abstractions: the Recursive Language Model (RLM) and the Continual Harness. With Opus 5, it achieves 95.5% on ARC-AGI-3, ...
010
Eric Florenzano @ericflo.bsky.social · 05/08/2026
See this is what I'm talking about www.theojaffee.com/p/levels-of-...
theojaffee.com
Levels of AGI-pilled
Making our definitions clear
000
Eric Florenzano @ericflo.bsky.social · 05/08/2026
VR is going to be so much fun when it swings back around again into the zeitgeist, and I think it'll be rich with creativity
010
Eric Florenzano @ericflo.bsky.social · 05/08/2026
No I'm talking about a belief beyond general usefulness. What prompted my post was this post by roon who believes in unbounded intelligence, which I think is nonsense, but it is a belief, so I can only sit and wait for it to bend into an s-curve like all else. Many beyond roon believe in this/RSI.
100
Eric Florenzano @ericflo.bsky.social · 05/08/2026
We need a proper noun for the religion of "being AGI-pilled" - and not one based around doom or optimism or de/acceleration, because those are different axes. Maybe singularitarianism, but that doesn't quite capture it or roll off the tongue.
230
Eric Florenzano @ericflo.bsky.social · 04/08/2026
Thank you!
011
Eric Florenzano @ericflo.bsky.social · 04/08/2026
I'm trying a thing where I do a developer stream on the weekend, and then edit it into videos to share on YouTube over the week. Here's a video of me yapping about new open models: Kimi K3, Inkling-Small, and DeepSeek V4 Flash 0731: youtu.be/58hb-rH9xzQ?...
youtu.be
Three Open Models Moved the Frontier: Kimi K3, Inkling-Small & DeepSeek V4
YouTube video by Eric Florenzano
010
Eric Florenzano @ericflo.bsky.social · 04/08/2026
A year later: no, yes, yes, no.
000
Eric Florenzano @ericflo.bsky.social · 01/08/2026
Actually not bad as a first line of defense, and it might help people understand that errors are an important steering surface even in non-adversarial agent/system co-design
030