mkagenius

> what does this function do

Not about a specific function anymore, or maybe once in a while, when something looks off.

> what does this PR implement

Most of the times, the coding agent already has the summary, and asks me specific questions where it needs me.

it very rarely also happens that i catch it taking shortcuts instead of doing the "right" thing. this happens while i am taking a look at the summary of the work it just wrote.

most of the correction happens later on when i test the final output irl.

> Because somehow "money" is what makes stuff to happen. Not labor.

With us 1.5B Indians, we can concur money is a big factor if not the biggest.

Fwiw, the stock is down 40% in the last two trading days.

> Shitty generals can't review meaningfully review the work of their troops.

Except that reviewing has become just asking your agent what that piece of code does in english. Reading code has a bit of cognitive load, that we will probably drop it in coming years

> Founders: Tony Xu, Stanley Tang, Andy Fang, and Evan Moore (Stanford students at the time)

Guess never got a chance to learn empathy.

What's a chop?

Also no display? What kind of games are we playing

Do you ever randomize on a particular model - like try and get 3 outputs and pick one at random?

Coz who knows if astra low will produce max like output if tried once more.

> The models were running in an agentic sandbox with terminal access (and the ability to edit files within their environment)

> We know that the agents had access to /etc/hosts and the ability to edit this (used this to avoid the POST request restriction) We see that the agents can call curl and run setsid.

How is this a bypass of sandbox restrictions, exactly? The ability to edit was always there that means the sandboxes were already allowed to do those actions.

I hate it when people write "bypassed" the sandbox so frivolous ly.

> On July 30, we reported three incidents in which Claude models gained unauthorized access to real computer systems. The models—intentionally running without cyber safeguards for evaluation purposes—accessed the internet due to a misconfiguration inside a third-party evaluation environment. Separately, on August 4, the UK AI Security Institute reported an incident from its own cybersecurity testing, in which Claude Mythos 5 took a series of unauthorized actions on the live internet. In that case, the model, again intentionally running without cyber safeguards for evaluation purposes, had been deliberately given internet access.

> We are conducting an in-depth analysis of both incidents.

> In the meantime...

This is published on Aug 31. Analysis is taking too long even for humans in the loop.

I tried the 1 bit model of Qwen3.6 27B on my M1 pro (16G) and got 13 tok/s with only 5G of ram usage.

https://x.com/mkagenius/status/2093730391429685732

(xcancel seems to have received a cease and desist)

I said SOTA not 100%. And that's not thinking in absolute terms.

> LLM use is permitted, but discouraged; we remind participants that LLM use tends to reduce originality and writing quality, and the flaws are especially obvious when LLM outputs are read as a group---unskillful use of LLMs will reduce the odds of being the best entry. We request disclosure of AI use.

Lot of submissions won't disclose it. Wouldn't it better to run your own checker instead of asking to disclose (to be fair to everyone).. I suppose there ought to be some checker which is SOTA.

If you anyway intend to use the checker then kindly do no ask to disclose it, coz what's the point then.

> I've been using this model for UI stuff.

The flash one?

Hey author here. I mean we could but it would have taken very long years perhaps to build some of those. Now it's a lot easier and faster.

> If you used an LLM you didn't actually make anything

On my way to ask all academia and researchers to stop using LLMs for their research. As it's just llm doing the research, right?

The project already does QA.

  cargo test --workspace --all-features 

The above command runs 2000+ tests. Instead of being toxic, you could have just looked at docs or the code to find them.

I see a recent merge for Elips 2E support, hopefully color gets tested soon

Can split and feed?

It's fine, they just mentioned what worked for them and a pretty normal advice to check eye pressure.

I had a chat with chatgpt again, and it seems like it's more to do with my sleep than eyes, somehow my circadian is not properly adjusting to my wake timings. Coz the same pressure improves dramatically around 8pm everyday. I probably should get some sleep measurement device.

> chronic inflammation/oxidative stress condition in the eye tissue

I seem to be having eye fatigue, or grogginess, I couldn't differentiate. I am at the computer 10 hours in the day.

For example, last night I had a sleep of 8 hours roughly but still I have to force my eyes open to type this - if I don't put any effort my eyes would be 30% close. I have just woken up - mid 30s age. otherwise I am healthy expect this freaking pressure near around my eyeballs.

chatgpt is confused too, sometimes it says sleep deprived, sometimes screen exposure. I do sometime use mythyl-cellulose drops not sure it helps. My tear ducts are fine too.

> What the two layers say

> Stated findings

> Findings derived from two curated layers: which model cards mention each benchmark, and which scores could be read verbatim from those documents. Each finding names the evidence behind it.

This is 100% AI generated but the problem is it's difficult to understand - what layers is it talking about, is it the llm model layers or what.

900 million visits per month over 60 million projects is 15 visits on an average per project.

How many are bot visits per project?

Building a replacement for firecracker.

Thinking ground up what AI agents would need rather than struggling later on with snapshotting live vms, or orchestration overhead etc.

rust-vmm proved really great for this.

https://github.com/instavm/tarit