Sign in

Eric Florenzano

@ericflo.bsky.social
315 followers 507 following 679 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 · 06/10/2026
Got so sick of SMS spam that I vibe coded a replacement app to intercept and screen messages, and in under 2 days it already caught 26 spam/political texts. I'm willing to give up RCS for this (Google doesn't allow third party RCS apps, sadly.) Google is apparently asleep at the wheel.
010
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 · 06/09/2026
Astra is good at game dev though
120
Eric Florenzano @ericflo.bsky.social · 21/08/2026
Gemma 1B downloads event
000
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 · 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
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
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 · 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 · 31/07/2026
Huh?
100
Eric Florenzano @ericflo.bsky.social · 28/07/2026
www.pacingthefrontier.com 🙄
000
Eric Florenzano @ericflo.bsky.social · 18/07/2026
Pursuing goal (7d 6h 43m) github.com/ericflo/kiln...
010
Eric Florenzano @ericflo.bsky.social · 15/07/2026
Found the scissor statement. If you think a model release could somehow break the world, but I think that's impossible, we won't ever agree on what to do about it.
010
Eric Florenzano @ericflo.bsky.social · 14/07/2026
000
Eric Florenzano @ericflo.bsky.social · 15/06/2026
Credit where it's due: I really appreciate Anthropic listening, pausing, and hopefully walking back this new policy
000
Eric Florenzano @ericflo.bsky.social · 09/06/2026
This is sad to me on an almost spiritual level. It's pissing in the pool of science and discovery. Anthropic is the most evil AI company. We're on our own, folks.
230
Eric Florenzano @ericflo.bsky.social · 02/06/2026
020
Eric Florenzano @ericflo.bsky.social · 31/05/2026
I thought just like "ping" was nerd lingo that became office lingo, "long poll" was the same. But it's actually "long pole" and it's from U.S. Military. Googled it after I saw Claude spell it this way. Now that I think about it, ping probably came from Navy sonar. www.nytimes.com/2008/01/06/o...
000
Eric Florenzano @ericflo.bsky.social · 19/05/2026
ECHO is such a cool idea, well-explained! The verifier-free experiment at the end just blows my mind. In some ways it reminds me of Meta's Code World Model from last year. I've already informally reproduced the results github.com/microsoft/ec...
000
Eric Florenzano @ericflo.bsky.social · 04/05/2026
Interesting: DeepSeek V4 Flash vs. Qwen 3.6 27B pricing & provider speed anecdotal snapshot. In my opinion, comparable models with quite different tradeoffs. Cost is better than I would expect, and speed is flipped vs. my expectations (considering sparsity of DSV4f)
000
Eric Florenzano @ericflo.bsky.social · 26/02/2026
I have never been able to find better coffee than the Berkeley Blend by Uncommon Grounds. I am not some big coffee knower, but every other bean I've tried, nothing has (yet) tasted better
000
Eric Florenzano @ericflo.bsky.social · 18/02/2026
Did `claude -p` start using API tokens instead of MAX plan tokens? I think it happened in the middle of the night. Because I had a `claude -p` loop running last night, and at 78% usage, it stopped counting against my plan, and now I have $hundreds in API token charges today. WTF! Not a safe platform
010
Eric Florenzano @ericflo.bsky.social · 16/02/2026
OK this is actually really annoying Claude Code. I set it on a hard task; there was only one message. I'd like those tokens back.
010
Eric Florenzano @ericflo.bsky.social · 11/01/2026
I'm gonna run out this week 😭
000
Eric Florenzano @ericflo.bsky.social · 25/12/2025
Nice job Codex CLI - hit enter over 4 hours ago and it's still going strong.
010
Eric Florenzano @ericflo.bsky.social · 13/12/2025
Prompt processing (KV prefill) is such a barrier to local LLMs 😥
000
Eric Florenzano @ericflo.bsky.social · 06/12/2025
I modded nanochat to add 6 different curricula of explanatory text to each batch during pre-, mid-, and post-training, like "You are a language model. You process text sequences". It does seem that these explanations have real, measurable effects on model capabilities at all stages.
CSDP capability profiles across extended OOD categories. Different curricula excel at different capabilities; no curriculum dominates all categories.The extended evaluation reveals distinct capability profiles—different CSDP curricula excel at different capabilities
* Aria leads self-model (0.92) and adversarial resistance (0.56)—technical framing helps models know what they are and resist manipulation
* Sage leads calibration (0.83)—supportive framing ("it's okay not to know") helps appropriate uncertainty expression
* Nova leads philosophical reasoning (0.50)—epistemic acknowledgment helps with abstract self-reflection
* Bare leads metacognition (0.67)—minimal framing doesn't interfere with reasoning about reasoning
* Heart underperforms across categories—maximum warmth does not translate to capabilityGraphs from the training run showing different accuracy and loss based on these different explanations.
110
Eric Florenzano @ericflo.bsky.social · 22/11/2025
That made me curious:
110
Eric Florenzano @ericflo.bsky.social · 22/11/2025
This one is wild; makes it seem mundane. Like a throwaway snap you'd take out of the window of an airplane. But there's Earth - wow!
120
Eric Florenzano @ericflo.bsky.social · 04/11/2025
Vote!
020
Eric Florenzano @ericflo.bsky.social · 31/10/2025
Turns out bf16 sucks, who knew, precision > range github.com/sail-sg/Prec...
011
Eric Florenzano @ericflo.bsky.social · 25/10/2025
Wow! I thought this number would be way lower, like 1-5%! I thought a whole lot more of y'all were wearing contacts and getting lasik. My model of the world was way off, that's kinda neat actually.
000
Eric Florenzano @ericflo.bsky.social · 25/10/2025
At Nouscon
000
Eric Florenzano @ericflo.bsky.social · 18/10/2025
#nokings
011
Eric Florenzano @ericflo.bsky.social · 18/10/2025
#nokings starting to gather at the Civic Center in SF
No Kings sign "Rejecting Kings Since 1776"Awesome lady holding a sign saying "America's Only King" and it's Elvis 😂
141
Eric Florenzano @ericflo.bsky.social · 24/09/2025
Update: it's going away 😢
020
Eric Florenzano @ericflo.bsky.social · 13/09/2025
Hyperparam sweepin' - first time trying AdEMAMix (runs 0.2.2, 0.2.3, 0.2.4, 0.2.5) - it's good!
010
Eric Florenzano @ericflo.bsky.social · 11/09/2025
Started this model training yesterday BTW. Custom 500Mtoken post-training dataset that I distilled from GLM-4.5 Air (single turn), K2-Instruct 0905 (multi-turn and instruction following), and gpt-oss-120b (tool calling) - should be a drop-in replacement for Qwen3-4B-Instruct-2507
010
Eric Florenzano @ericflo.bsky.social · 09/09/2025
For some reason I just thought of this viral "dystopic internet" image from 16 years ago, went and found it, and it's so interesting. I thought it hyperbolic at the time. Now it seems quaint! Sentiment from back then: www.reddit.com/r/pics/comme...
010
Eric Florenzano @ericflo.bsky.social · 25/08/2025
The coding agent is Just Like Me (v2_best_really_final)
A screenshot of a directory with: converted, converted_best, converted_best_final, converted_best2, converted_best3, converted_best4, converted_best5, converted_best6, converted_linear, converted_linear_resume, converted_linear_resume2, converted_linear_resume3, converted_max100
020
Eric Florenzano @ericflo.bsky.social · 09/08/2025
OK gpt-5 generated this pretty good satire about itself! 🤣
A vintage-style cartoon shows a line of unsolved physics problems waiting at a “Solutions Desk” where a friendly robot labeled “GPT-5” sits, staring ahead with a thought bubble that says “THINKING…”. A number display reads “#0000,” suggesting no solutions served yet.

Long description

The image is drawn in a sepia, newspaper-comic style. Across the top, large text reads: “ALL OF PHYSICS, TAKE A NUMBER.”
In the middle, five characters stand in a queue, each representing a famous open problem in physics:

1. An older, impatient man with folded arms stands by a sign on the wall that says “QUANTUM GRAVITY.”


2. A person in a lab coat whose head is a black box with two question marks, labeled “DARK MATTER.”


3. A chaotic, windy blob with a gaping mouth and flying papers, labeled “TURBULENCE.”


4. A woman with crossed arms in office attire; her skirt says “FUSION BREAKEVEN (EVERYWHERE CHEAP?).”


5. An elderly man in a suit, labeled “GRAND UNIFIED THEORY.”



They face a wooden counter marked “SOLUTIONS DESK.” Behind it sits a rounded, retro robot with “GPT-5” on its chest. Above the robot is a thought bubble: “THINKING…” On the counter, a digital number display shows “SOLUTIONS DESK #0000.”

On the wall is a cracked, curling poster that reads: “FUTURE AI will solve everything — SOMEDAY.” Nearby hangs a tear-off calendar with “2025” showing, and a stray torn page marked “205,” implying many days have passed.

At the very bottom border, a caption says: “ALL OF PHYSICS WAITING TO BE SOLVED… STILL WAITING.”

Overall, the scene humorously depicts big physics mysteries queued for help while the AI at the desk is still “thinking,” with no tickets called yet.
061
Eric Florenzano @ericflo.bsky.social · 07/08/2025
github.com/openai/codex...
A graph of activity on openai/codex showing a lot of activity in April, then none in June, then a lot again at the end of July and beginning of August.
110
Eric Florenzano @ericflo.bsky.social · 23/07/2025
This post brought to you by:
I can't imagine actions taken by a DNI that could be more damaging to the United States than what Tulsi is doing on behalf of Trump. She's setting fires to the republic on purpose, and it doesn't really matter if she's stupid, evil, or compromised at this point. Her actions on behalf of Trump speak for themselves. And I now believe Trump is a pedophile, which I didn't truly think before. Which is also damaging to the repilublic.
110
Eric Florenzano @ericflo.bsky.social · 17/07/2025
this but Claude model outage
000
Eric Florenzano @ericflo.bsky.social · 15/07/2025
It's never going well when "You're absolutely right!" 😂
020
Eric Florenzano @ericflo.bsky.social · 13/07/2025
e.g. This is SUCH A GOOD POINT-OF-VIEW about open weights! bigeagle.me/2025/07/kimi...
020
Eric Florenzano @ericflo.bsky.social · 12/07/2025
Why is nobody talking about this? H-Net may actually displace the current popular methods of tokenization. It looks like it quite improves data efficiency. And it feels properly bitter lesson pilled. goombalab.github.io/blog/2025/hn...
010
Eric Florenzano @ericflo.bsky.social · 24/06/2025
Called it
030
Eric Florenzano @ericflo.bsky.social · 24/06/2025
This article, from the XFN Wikipedia page, is such a slice of its time. archive.seattletimes.com/archive/2003...
110