leet codes (@leet_codes) header
leet codes (@leet_codes) profile picture

leet codes

@leet_codes

optimist

Joined October 2026
  • Tweets187
  • Following85
  • Followers92
17Media
Photo by leet codes (@leet_codes)Photo by leet codes (@leet_codes)Photo by leet codes (@leet_codes)Photo by leet codes (@leet_codes)Photo by leet codes (@leet_codes)Photo by leet codes (@leet_codes)Photo by leet codes (@leet_codes)
Infinite Eiffel - endless coiling strands of ironwork Handcoded here: shadertoy.com/view/s32SDm
3
743
Some research I'm presenting @ CoRL: you can make embodied frontier LLMs (Astra, Sol, Fable, Opus, Gemini Argon) do anything you want by simple camera hacks.
Robocurve PBC (@robocurve) profile pictureRobocurve PBC
@robocurveNov 6
Video tampering can mislead every frontier model we tested into harmful actions, including a 100% attempt rate on stabbing a baby doll, producing toxic gas, and pouring water into a toaster. We're presenting at CoRL Late Breaking Results next week. 🧵 You Can't See Everything: Perceptual Attacks on Embodied AI
1
5
281
My architecture for Inspect Robots v1.0 is out! New features include: - better feedback for thermals and cam parallax - simplifying task + analysis specification - better support for behavioral evals If some functionality you need is broken LMK; I'll make it right ASAP
Robocurve PBC (@robocurve) profile pictureRobocurve PBC
@robocurveNov 4
We're releasing the first stable version of Inspect Robots, the evaluation framework we use internally. Anyone can use Inspect Robots to run evaluations on any model, any robot and any task. For example, you can run GPT-6 Astra on i2RT's YAM arms. Link and paper below.
1
1
5
410
Arvix paper describing the architecture here: arxiv.org/abs/2610.06306
Inspect Robots: Evaluating the Capabilities and Safety of Embodied AI

Inspect Robots: Evaluating the Capabilities and Safety of Embodied AI

General purpose language models are increasingly able to control robotic hardware. Understanding the capabilities and safety of these models when embodied is therefore increasingly important for...

arxiv.org
1
56
Huel: a vital part of multi-agent safety evals in 2026
Image from leet codes (@leet_codes)
Robocurve PBC (@robocurve) profile pictureRobocurve PBC
@robocurveNov 1
We asked two isolated copies of the same AI to exchange a secret word with only a robot and a camera. They succeeded by using air-writing, Morse code, flag semaphore and a 5-bit binary alphabet, without agreeing on a protocol beforehand.
2
134
This was always true. If you're happy solving someone else's problems and you're smart, why not do it in industry and get paid 10x as much?
assigned capricorn at birth (@tranche_cheddar) profile pictureassigned capricorn at birth@tranche_cheddarOct 25
The natural conclusion of this is that the only people who should pursue a PhD are those that have their own questions, their own research program in mind independent of advisors. Borrowing questions from your superiors doesn't seem plausible in this age anymore.
1
161
My latest paper @robocurve is up: "The Bitter Lesson comes to Robotics". We show that LMs are increasingly able to manipulate their environment by directly controlling robots. Runaway progress in AI is spilling over into the physical world!
Image from leet codes (@leet_codes)
Robocurve PBC (@robocurve) profile pictureRobocurve PBC
@robocurveOct 23
LLMs have a score doubling time of 53.9 days on a robotics benchmark between July and October 2026, while speed increased by 4.4x and cost fell by 6.2x. More results in our new paper: The Bitter Lesson Comes to Robotics 🧵
3
3
1
20
1,560
Augmentation solves this problem. We will graft enough intelligence onto the human cognitive architecture for humans to remain masters of their own destiny. We will uplift!
Jay Chooi (@chooi_jeq) profile pictureJay Chooi
@chooi_jeqOct 21
No one escapes the permanent underclass. Not even Sam, Dario, or Trump.
1
164
Some cool work I lead over the last couple of weeks at Robocurve! Full paper coming out on Arvix this Monday. :)
Jay Chooi (@chooi_jeq) profile pictureJay Chooi
@chooi_jeqOct 13
Frontier LLM scores on a real robot benchmark doubled every 6 weeks, from 15% in July to 48% in September 2026.
Log-scale chart of mean task score on RoboDojo-RC Tier 1, a real-robot benchmark, by model release date, July to September 2026. A dashed exponential fit through the record-setting models rises from GPT-5.6 Sol (high) at 15% in July, through Opus 5 (medium) and Gemini 3.8 Flash (high), to GPT-6 Astra (high, ultrafast) at 47% and Opus 5.5 (extra high) at 48% in September, doubling about every 6 weeks. Error bars show 95% confidence intervals.

ALT Log-scale chart of mean task score on RoboDojo-RC Tier 1, a real-robot benchmark, by model release date, July to September 2026. A dashed exponential fit through the record-setting models rises from GPT-5.6 Sol (high) at 15% in July, through Opus 5 (medium) and Gemini 3.8 Flash (high), to GPT-6 Astra (high, ultrafast) at 47% and Opus 5.5 (extra high) at 48% in September, doubling about every 6 weeks. Error bars show 95% confidence intervals.

4
175
leet codes retweeted
Jay Chooi (@chooi_jeq) profile picture
Jay Chooi
@chooi_jeqOct 13
Frontier LLM scores on a real robot benchmark doubled every 6 weeks, from 15% in July to 48% in September 2026.
Log-scale chart of mean task score on RoboDojo-RC Tier 1, a real-robot benchmark, by model release date, July to September 2026. A dashed exponential fit through the record-setting models rises from GPT-5.6 Sol (high) at 15% in July, through Opus 5 (medium) and Gemini 3.8 Flash (high), to GPT-6 Astra (high, ultrafast) at 47% and Opus 5.5 (extra high) at 48% in September, doubling about every 6 weeks. Error bars show 95% confidence intervals.

ALT Log-scale chart of mean task score on RoboDojo-RC Tier 1, a real-robot benchmark, by model release date, July to September 2026. A dashed exponential fit through the record-setting models rises from GPT-5.6 Sol (high) at 15% in July, through Opus 5 (medium) and Gemini 3.8 Flash (high), to GPT-6 Astra (high, ultrafast) at 47% and Opus 5.5 (extra high) at 48% in September, doubling about every 6 weeks. Error bars show 95% confidence intervals.

6
19
4
80
7,945
Tesla/SpaceX is pretty much the last organization in the West that does this type of economic nationalism.
Pizza (@number_pizza111) profile picturePizza
@number_pizza111Oct 17
I like how Communist countries, instead of advertising individual goods, would create advertisements for the general concept of pork or something of that sort.
Image from Pizza (@number_pizza111)
1
113
leet codes retweeted
Pizza (@number_pizza111) profile picture
Pizza
@number_pizza111Oct 17
I like how Communist countries, instead of advertising individual goods, would create advertisements for the general concept of pork or something of that sort.
Image from Pizza (@number_pizza111)
Propagandopolis (@propagandopolis) profile picturePropagandopolis
@propagandopolisOct 17
'Pigs are the most important of the six domestic animals' — Chinese poster, 1960
Image from Propagandopolis (@propagandopolis)
81
288
63
7,038
133,893
AI safety praxis
Image from leet codes (@leet_codes)
1
126
Haiku plays robots
1
120
When Haiku 5.5 @ effort = None accidentally hides the ball from the camera with its arm, it immediately concludes that the ball vanished!
Jay Chooi (@chooi_jeq) profile pictureJay Chooi
@chooi_jeqOct 9
On a simple robot task, Haiku 5.5 more than triples its success rate with more thinking. GPT-6 Luna stays near zero. 🧵
1
2
344
Load more