Ashutosh Mathore

founder @tryatlensa
building at the intersection of ai and energy

[Ashutosh Mathore @ashutoshftw]

Had jev, claude haiku 4.5, and gemini 3.5 flash-lite play tetris. Same 200 pieces, same legal moves, real time.

Jev: 9200 pts, 75 lines, 300ms/move, 0 errors
Claude: 8900 pts, 72 lines, 1.52s/move, 0 errors, $0.48
Gemini: 9000 pts, 75 lines, 1.13s/move, 0 errors, $0.02

Jev wins on score and is doing it 4-5x faster per move than either LLM, for basically free. That gap isn't the model being smarter, it is the model doing less work per decision

[Carter Church @CarterWChurch]

I used GPT-6 Astra to break an unsolved cipher to one of Napoleon's generals that had gone unread for 217 years.

What makes this impressive isn't actually the codebreaking, but that Astra completed the entire multi-modal workflow in ~6 hours from a single image and goal.

1/6🧵

[Ashutosh Mathore @ashutoshftw]

Ran into some really cool people at DevDay

[Ashutosh Mathore @ashutoshftw]

@SpenzDigital yep, that’s exactly the kind of agent behavior I want to catch

[Ashutosh Mathore @ashutoshftw]

Jev unlocked a new way to stop AI agents from spinning.

Agents do not always repeat the same action. Sometimes they try different things based on the same wrong assumption.

I built a semantic circuit breaker around that.

In one stuck run:
12 model calls -> 4
19.9K tokens -> 5.1K

https://t.co/fR44bfNDrA

[Ashutosh Mathore @ashutoshftw]

@catmanyau Exactly. Different actions can still be the same failure mode underneath

[Ashutosh Mathore @ashutoshftw]

@laichunpongben I think they’re a bit overkill for this benchmark. The point wasn’t max intelligence. Haiku and Flash-Lite are already good enough at the task. I was testing decision speed, so I intentionally picked the fastest Claude and Gemini models I could find.

[Ashutosh Mathore @ashutoshftw]

Had jev, claude haiku 4.5, and gemini 3.5 flash-lite play tetris. Same 200 pieces, same legal moves, real time.

Jev: 9200 pts, 75 lines, 300ms/move, 0 errors
Claude: 8900 pts, 72 lines, 1.52s/move, 0 errors, $0.48
Gemini: 9000 pts, 75 lines, 1.13s/move, 0 errors, $0.02

Jev wins on score and is doing it 4-5x faster per move than either LLM, for basically free. That gap isn't the model being smarter, it is the model doing less work per decision

[Ashutosh Mathore @ashutoshftw]

Code: https://t.co/XPl9aUsyHL

npm: https://t.co/ei5ElBAFlf

npm install progressgate

Built with early access to TypeSafe Jev.

[Ashutosh Mathore @ashutoshftw]

Example failure mode:

check logs
↳ permission error
try sudo
↳ same blocker
change path
↳ same blocker
switch account
↳ same blocker

Different actions. Same broken assumption.

[Ashutosh Mathore @ashutoshftw]

Some use cases Jev unlocked or rather made it practical at scale:

- Optimized power grids
- Smarter robots
- Autonomous vehicles getting more safe
- Inspecting infrastructure with drones
- Faster scientific discovery
- Social media algorithms more addictive
- Smarter real-time trading
- Better anti-cheat in games/casinos

[Ashutosh Mathore @ashutoshftw]

Final scoreboard from that run. Jev takes it on score with a fraction of the latency and basically no cost

[Ashutosh Mathore @ashutoshftw]

Interesting way to keep North Koreans out of US remote jobs lol

[etn. @etnshow]

JUST IN: NVIDIA's $12,930,300,000 acquisition of Hugging Face contains an easter egg. The number 129,303 is the decimal conversion of Unicode point U+1F917.

The 🤗 emoji.

[Ashutosh Mathore @ashutoshftw]

Disappointed with the naming

[OpenAI @OpenAI]

This is GPT-6 Astra.

Anything you can do on a computer, Astra can do for you. Fast.

[Ashutosh Mathore @ashutoshftw]

@shl Yessir!

[Ashutosh Mathore @ashutoshftw]

Let's goooo!!! super excited @OpenAI

[Ashutosh Mathore @ashutoshftw]

Spent the weekend writing direct API clients for every US electricity market because I didn't want a dependency

Turns out CAISO, ERCOT, NYISO, MISO, SPP, ISONE, and PJM all have public endpoints. none of them need an API key for the data I needed.

So I packaged it up:

```
from kardashev import CAISO, NYISO

caiso = CAISO()
caiso.get_fuel_mix()
caiso.get_lmp(market="RT", node="TH_NP15_GEN-APND")
```

`pip install kardashev`

Using it to power a live nodal LMP map, carbon intensity tracker, curtailment data, and interconnection queues - all pulling straight from the source

https://t.co/DEbohPF8um if you work with energy data