Robert

AI | Startups | Thinking out loud

[Robert @robertirv1]

In a reward model I found you could turn off 90% the activations in the second half of the NN with no loss in performance.

Just need to keep the final N position activations and pass through the residual for all positions greater than N.

[Robert @robertirv1]

AI is not better than humans at mathematics (yet). It’s better at a subset of it. In the same sense AI is better at a subset of programming.

The fact that many of the solutions are similar like “we improved the bound to 0.9998” tell you this.

[Robert @robertirv1]

@corsaren AI is good at solving a subset of maths problems that look messy like this. It doesn’t prove the actual true bound is not elegant.

[Robert @robertirv1]

@alz_zyd_ AI as it stands is probably really good at solving a particular subset of mathematics problems. But is weak at solving other problems.

[⚡️NEVERDOOMPRO⚡️ @NEVERDOOMPRO]

NEVER DOOM.

[lusso @luusssso]

The incredible colors of 1980s Japanese nightclubs

[Arbutus Tree @aphercotropist]

Speedrunners literally decompiled The Legend of Zelda: Ocarina of Time, which took two years, in order to better figure out how to manipulate game memory through bugs, which led to an arbitrary code execution technique and a sub 10 minute world record, two decades after release

[Robert @robertirv1]

Happy Independence Day.

[Robert @robertirv1]

GLM 5.2 achieved its gains by being more aggressive with RL. I can see this by looking at word overuse index on my prompts.

It resembles Gemma 4 in this regard which is another model that was aggressively RL’ed and this made it difficult to find tune for other tasks.

[Robert @robertirv1]

AI has improved my code quality not just speed.

Because I can quickly do refactors. And avoid taking shortcuts.

Just requires some care like doing one small commit at a time and keeping high standards.

[Jedi @dhe3g]

Have you ever stopped to think about how everything used to be created to generate peace of mind and accept those waiting periods?

[Barrett @SledgeDev]

How it felt to be a software developer before ai.

[Robert @robertirv1]

The performance of an LLM as a judge is a function of compute just like many other AI task.

Thinking longer makes a better judge. Best of N. Larger models. Judging one example at a time rather than a batch.

[lusso @luusssso]

When was the last time you saw a corporate space of any kind that blew you away

Time to take chances again

[Robert @robertirv1]

@PaulSkallas It’s kind of like grip strength. It’s the first thing to go.

[Robert @robertirv1]

I viewed my AI work as being a “scientist” which meant an emphasis on understanding. It’s going very poorly this year.

I think a better model is “esports”. Emphasis is on clicks per minute, volume of experiments, experiment strategy, psychology, racing to the goal.