Design Engineer, Artist, & Researcher โ Infinite Play Wins
San Diego, CA
- Tweets 13,990
- Following 5,322
- Followers 1,908
- Likes 35,917
Pinned Tweet
Finally feel like I have a website that reflects my design work and philosophy. This has been a labor of love, go check out the fish and try asking some questions at the link below.
6
2
13
1,444
I had to roll my own recognizer for MetaMedium and my site, it does basic polygons and lines/arrows (& custom). which is why Jev is so exciting for me, as recognizing more complex shapes this fast will be a huge boost. multimodal will be the final unlock. jjh111.github.io/MetaMedium/
Jev as a shape recognizer.
Results are mixed, but damn itโs fast.
Multi modal coming though (See comments), hopefully that โjust worksโ
1
5
472
My site search is an example of why Iโm excited about the Jev era: I was already offering as much as possible in structured Json and clear categories with LLM as extra rather than dependency. Itโs like we had the cake (BM25, etc) and the cherry (LLM) but now we have the Icing ;)
Design is structured play, and I'm the disciplined jester.
johnhanacek.com
1
122
My personal site is like a demo-ground for my true ambition to draw only living fish lol jjh111.github.io/MetaMedium/
32
I was literally giving up on doing my full system 0, 1 & 2 paradigm of watching the canvas live b/c I didnโt have a good model for complex library matches -0.8b qwen is too slow still & too big for a deterministic lookup. Then we got Jevโd & turns out system 1 is the new hotness
120
This right here, when I first read about BPL back in 2015 it seemed so great: structured fuzzy decision making. The perfect middle ground between computer code and heuristics. Work on Meta medium has shown me that thereโs still a gap between smart low-level library look up and Full LLM generative intelligence. Then here comes Jev offering to bridge that gap and make my dreams come true. @typesafeai
i'm excited about Jev for generative interfaces because if you can remember back before transformers were invented, we did generative interfaces with decision trees & constraint solvers. it wasn't fully deterministicโi think having probability over fixed answer is GOODโbut things were more repeatable
of course, LLMs were kind of better at learning and parsing intent than anything before them, but no matter how good LLMs get at generating interfaces, they're always black boxes that are sometimes magical, sometimes abysmal, and always Incredibly Frustrating To Nudge As A Designer
I actually yearn for more control over my systemized design environments than the blackbox Slop Or Not randomness that LLMs inherently provide
anyway, i am extremely excited about the new possibilities Jev is opening up in Determinishtic Interfaces, and in intelligent software composition in general
1
190
And itโs not the only option even if it is the newest. Will be experimenting more.
Context compaction with GLiNER2.5. โจ
Completely local (open weight) with span extraction for verbatim context preservation.
For each completed tool interaction, GLiNER2.5 sees:
โข Current goal
โข Nearby conversation
โข Tool name and arguments
โข Original result
It then chooses to:
โข Keep the full interaction
โข Keep exact evidence
โข Keep only the rerunnable call
โข Drop an irrelevant or superseded pair
Deterministic safeguards preserve recent messages and mutations, keep uncertain interactions in full, reject invalid spans, and maintain tool-call/result pairing.
When a 96,000-character log contains only a few useful failure lines, the compactor copies those exact spans and removes the surrounding noise. User and assistant messages remain unchanged.
Everything runs in a local Python worker after the checkpoint download. Before replacing history, it validates offsets, confidence values, conversation text, and tool pairing. Failures fall back atomically.
Shadow mode is enabled by default so you can inspect the decisions first.
Code below :)
26
Design is structured play, and I'm the disciplined jester.
johnhanacek.com
5
357
See this is part of why I am working so hard on a living metaphorical canvas substrate, the idea would be it is morphing out its thinking and work as river, rocks, terrain metaphors that each invite different levels of โeditingโ at different time scales. Less โharnessโ more ecosystem of the visual state of mind shared as a picture evolving over timeโฆ jjh111.github.io/MetaMedium/โฆ
the most frustrating thing about longggg running agents is being a babysitter on a million threads in the evening wondering when theyโll be done and i can go home (or taking my laptop with me and hoping my iPhone tethering survives the commute)
i have a much more involved Humane Control Plane from this spring iโm porting to bb, but my stopgap was to make an automation that just updates one thread when my agents & evals finish a task or need my input after 5pm
this + seamless remote access lets me keep easy tabs on them without being On My Laptop Going Through A Million Sessions All Night
(i know codex and others have automations but i just never vibed with them idk, bb feels different bc the rest is so malleable and i can see the universe i wanna create with it)
((bb is really good and you should use it))
(((still working on remote sandboxes also but honestly itโs nice just knowing that my laptop is plugged in and chugging away at my desk and i have literally the very same UI being served to my phone & other devices)))
1
1
3
199
Designer for > 10 years I gotta be honest: Now that I have One Black Box to Rule Them All of LLM code gen, I find it hard to justify using any platform when I can just code. Flashbacks to XD & sketch & figma lock in! so Elyx & dozen other similar ideas, Iโm still skeptical.
1
2
171
In light of Jev & the general coming of age of probabilistic computing, I wanted to share an old class blog from 2015 @GeorgetownCCT (typos & all! human provenance) where I first discovered Lake, Salakutdinov & Tenenbaum's Bayesian program learning (BPL) johnhanacek.com/writing.htmlโฆ
1
3
119
Ok this could be the infinity stone I needed. I'm having to fight LLMs into structured outputs, so this is very promising as part of the mix of experts for the magic canvas. jjh111.github.io/MetaMedium/
The gains arenโt free: Jev can't generate text
Comparing Jev vs LLMs side-by-side makes the trade-off clear
Fun fact: replacing sequential computation with parallel is the same way Transformers leapfrogged RNNs
1
218
upgraded my Nous Portal subscription to get more work done, having more fun anyway @NousResearch
1
1
158
piped.video/NdSD07U5uBs?is=pvnaโฆ
Slide from 35min mark but whole talk is amazing about how things really get made from Alan Kay himself
isn't the whole point of research labs that you can move slowly and develop long-term creative ideas? do we think at PARC they were like "you have to invent the next 50 years of computing by EOD tomorrow"?
1
2
173
Dear tpot, I need your best Claude/agent management memes for an upcoming course presentation. I keep trying to find the drawn one thatโs based on Daniel Radcliffe walking the dogs but itโs all the Claudeโs. Any other good ones?? Letโs collect the greatest hits! W/ credits thnx
1
116
Me when I'm working on three simultaneous projects but hit the Fable 5.1 weekly limit and don't have budget for Astra right now while my GLM 5.3Flash and Opus 5 sessions are getting to the more complex parts of the pre-made dev plan, yet I can't work with Director Fable to review their work until Saturday...
1
1
82