Trent Zock-Robbins

• 100% bitter-lesson-pilled
• localmaxxing Qwen 27B on 4 × 3090
• → byom llm chat on iOS
• veteran swe, ex-chemist, touching grace

[Trent Zock-Robbins @tazr_dev]

Flatmachines 4.2.0

• native state machine snapshot auto-prune
• fail fast on missing profile file
• migrate tasks to tk: memgrafter/tickets-rs
• log to stderr by default for headful cases

This release was written with deepseek-v4-flash and qwen-3.6

https://t.co/D7ANWxeOS7

[Trent Zock-Robbins @tazr_dev]

@ShimazuSystems That moat actually lasted negative 1.5 years.

I believe Laya is 1.5 years old, it's just a different marketing tack due to perceived PMF.

• https://t.co/2wz8I6WUMW
• https://t.co/XsgGPUYZao

[Trent Zock-Robbins @tazr_dev]

Foveated reasoning is really something.

https://t.co/gXaM12EFsH

[pomterre @pomterree]

What if you scaled DJev wrt. reasoning at decode time?

I added a 512-token budget to the think knob (was 0 by default); where the model now reasons before returning a structured read.

A tuned 26B diffusion model can now trivially beat Jev! 🙃

[Trent Zock-Robbins @tazr_dev]

@urchadeDS @fastinoAI Of course, I linked two posts working with doom in Gliner. I needed to build a map (or any programmatic or intelligence interface) of GLiNER2.5 current state to my desired action.

[Trent Zock-Robbins @tazr_dev]

Other notable improvements:

• Optimized embedder for reuse: 11.6ms/embed
• Total decision time: 15ms or 65hz
• Peak memory: 1.9 GB

You read that right, 77% of decision time is embedding!

I'll want to optimize embedder speed & format next.

Can I get to 500hz decision rate?

[Trent Zock-Robbins @tazr_dev]

Yesterday, GLiNER2.5 played doom. Rapid play looks coherent, but Jev has a secret sauce: decisions.

So, I kept the embedder and trained a small decision head on the output.

Now it's a homebrew decider model on embeddings.

What homebrew decider model will you make this weekend?

[Trent Zock-Robbins @tazr_dev]

GLiNER2.5 Doom, code tomorrow.

• mac mini m4 16gb
• 10.9 decisions/s
• 34 tics/s

[Trent Zock-Robbins @tazr_dev]

@Stellanhaglund Just added below main post, thanks for asking!

[Trent Zock-Robbins @tazr_dev]

Code, as promised. This is the latest trainer iteration, it doesn't include the initial GLiNER2.5 to decision mapping code.

https://t.co/xay8C81iG1

[Trent Zock-Robbins @tazr_dev]

@ChuanxinL It is hard to recommend one yet, it comes down to benchmarking for your project. Jev is managed, but closed. It looks promising, but hard to rely on a closed API for core infra.

[Trent Zock-Robbins @tazr_dev]

GLiNER2.5 Doom, code tomorrow.

• mac mini m4 16gb
• 10.9 decisions/s
• 34 tics/s

[Nathan Wilbanks @NathanWilbanks_]

@george_onx using a GLiNER like model to play doom and do a bunch of business tasks

works great, its like having open source Jev at home!

https://t.co/hH91b6Tqf7

[Trent Zock-Robbins @tazr_dev]

@EmilevanKrieken It is based on the Jev doom demo, which also uses text. Currently I am unbinding some of the text component.

https://t.co/0WH7DpySfH

[Trent Zock-Robbins @tazr_dev]

Dunking on Jev is going to be replaced VERY quickly by using Jev... Already is.

Hope you guys are ready for the claws.

https://t.co/cxeb4sNUHF

[Gregor Zunic @gregpr07]

Breaking: Browser Use + Jev = Ultrafast ⚡

Findings flights took 7s and cost only $0.0039 🤯

> new action space every step
> DOM state space
> small LLM fallback to type

(this video is at 1x speed btw)

Built a tiny open source browser agent. try it below ↓

[Trent Zock-Robbins @tazr_dev]

@CompleteSkeptic @badlogicgames Love the focus on cognitive core!

[Trent Zock-Robbins @tazr_dev]

@george_onx @NathanWilbanks_ Here you go! https://t.co/EGQAefCZYe

[Trent Zock-Robbins @tazr_dev]

GLiNER2.5 Doom, code tomorrow.

• mac mini m4 16gb
• 10.9 decisions/s
• 34 tics/s

[Trent Zock-Robbins @tazr_dev]

@mmastrac Brutally dunked on by the bitter lesson. I feel pain just reading this.

[George Maloney @george_onx]

We've been reading the comments. Weights are on Hugging Face.
https://t.co/6khtZDUNOJ

[Nathan Wilbanks @NathanWilbanks_]

@george_onx using a GLiNER like model to play doom and do a bunch of business tasks

works great, its like having open source Jev at home!

https://t.co/hH91b6Tqf7

[Rohan Paul @rohanpaul_ai]

Jensen Huang pushed back hard against AI labs calling for tighter regulation and slower AI development, at Dreamforce.

"If you are not confident about the safety of the products, and you're not confident in its functionality, capability, or safety, then don't just release it.

We don't need any new laws. We don't need new regulations. we just need companies to decide that when it's time to run as fast as they can."

safety and speed are not mutually exclusive, and new laws are not required to make AI systems safer. Build fast, test aggressively, and pause only when you are no longer confident the product is safe.

---
From "Salesforce" YouTube channel, (full video link in comment)

[Trent Zock-Robbins @tazr_dev]

@bnjmn_marie I like the idea of cost-based tries! Basically give the model a meter.