inactive

@KingBootoshi @Juicetin You can have them loop usefully if you cap the number of rounds and make sure you’re resuming the session on each round so it’s the same instance of Claude and Codex looping and not fresh bots starting new threads each time

@Juicetin do they get into recursive loops or can they actually agree on a base point ???

you guys know you can throw more compute at a problem yourself right? if one agent couldn't solve it, throw a diff agent at it who can then see the proposed solutions and offer a variety of different ones BIG help especially when they're different LLM models works everytime!

@systematicls it’s conscious of its own context window for sure. it’ll rush to land a slice of work before compacting, even if that means faking a bunch of stuff to get there.

Claude literally sees session length as a legitimate reason to end the session, and that is preventing it from doing its best work because it becomes obsessed with ending the session. This is so pervasive I'm almost certain it's RL'ed behavior or written into the harness of CC.

@doodlestein @karpathy @Qivshi1 needs to be mediated by a significance factor or something “it’s a bad idea to touch a hot stove” is the kind of memory that should persist without reinforcement

@karpathy @Qivshi1 Seems like an easy fix with a time-decaying salience factor, kind of like synaptic strength fading when a pathway isn’t reinforced.

One common issue with personalization in all LLMs is how distracting memory seems to be for the models. A single question from 2 months ago about some topic can keep coming up as some kind of a deep interest of mine with undue mentions in perpetuity. Some kind of trying too hard.

@mkurman88 @jake_researcher yeah claude lies now

Codex? Claude was the faker. It wrote tests to pass, not tests to verify, and that's the point. Codex took a different direction. First, created a test based on the problem description; then, it found the cause and fixed it. Then iterated on several possible variations and fixed them too

I just canceled my Claude subscription. It couldn't resolve a straightforward issue for 3 hours, even though I pointed out several times where the bug lies (I was afk, so I couldn't do it by myself, but that's not the point). I asked Codex to fix it -> 1 minute and done. I asked Codex to review other things Claude made - with GSD, 20 steps, discussion, plan with research, execution. It has been working since Saturday. And what? Everything was faked. I don't get it...

@justalexoki i think for eric it’s just regular psychosis

worst ai psychosis i've seen so far

@SoilOfHarvest @jvgraz @RWApodcast the part where that doesn't matter because they distribute command in advance

@jvgraz @RWApodcast What part of this chess game was having the entirety of your leadership killed on day one?

One advantage the Iranians have is that their entire elite seems to consist of 130 IQ polymath combat veterans with PhDs who stormed across minefields with fixed bayonets surrounded by poison gas back in the day and then spent 30 years studying philosophy, game theory, engineering and military science. Average decision-maker quality appears to be dramatically higher than elsewhere, especially qualified for shit-hits-the-fan situations

@BenReynen @SandRepeat @PerryALPHA we are well beyond renaming individual streets

@SandRepeat @PerryALPHA Same thing in Canada but we have taken it a step further and changed the names of streets to indigenous language. People living on the street are unable to change their driver’s license because the characters don’t exist in the application form. Don’t ask me how to pronounce it.

The aboriginal translator

@sterlingcrispin Gemini will pick up bugs that Codex and Claude both miss. It’s worth having all three in the mix.

Claude 4.6 is a good programmer but writes insanely severe bugs constantly, it won't catch them all in audits, nor will other claudes You need codex 5.4 auditing every commit 4+ times. If you don't believe me, try it. I have an /auditcodex skill for it github.com/sterlingcrispi…

@_ggLAB @SamMonsonNFL meditation in this context means "a written or spoken discourse expressing considered thoughts on a subject" which is typically something that requires introspection

@Auinout @SamMonsonNFL no not the book but the act of meditation

What is the motivation to keep doubling and tripling down on such an obviously stupid take? Even if you think your interpretation of that quote is correct - and there's a good chance it's not... The fucking book this is taken from is called 'Meditations' you dipshit.

@nate_yiu @odysseus0z It's referenced in the readme on the Symphony Github too github.com/openai/symphon…

@nate_yiu @odysseus0z Sure. You can just point your agent at it and have it apply the principles and practices: openai.com/index/harness-…

@rishdotuk @_julianschiavo @giffmana I found them for anyone following along at home: github.com/lucasb-eyer/do…

@rishdotuk @_julianschiavo @giffmana dm'd

It's so frustrating how GPT-5.4 sprints away and writes tons of code that massively overcomplicates a simple bug fix or UI tweak :(

@_ggLAB @SamMonsonNFL How does the book arrive at its conclusions?

@SamMonsonNFL Meditation has nothing to do with introspective thinking. The purpose of meditation is to free yourself from introspective thinking you dipshit. Read between the lines. The message is : Just build. If you have time to whine about your feelings, just build.

@pmarca the book is literally called meditations and talks almost exclusively about being a good man

Marcus Aurelius, Meditations, X.16: “To stop talking about what the good man is like, and just be one.”

@Linahuaa the worst part is he pulled an introspective marcus aurelius quote from a book literally called meditations which mostly talks about being a good man

Dude crashed out, went to sleep, woke up, and went right back to crashing out. He's spending the whole day researching for clever-sounding comebacks, instead of listening to his own advice of not ruminating, just taking the L, and moving forward with being a 10x Marcus Aurelius engineer chad or whatever he sees himself as. I blame LLM oneshot psychosis. Garry Tan is showing the same symptoms...

@Aizkmusic A lot of metal falls into this category because the lyrics are indecipherable and don't register as words. Also Tom Waits.

I have a theory that people who listen to instrumental music or music in languages they don't understand have very loud and chatty internal monologues.

@rishdotuk @_julianschiavo @giffmana Where can I find these?

@_julianschiavo Use @giffmana’s deslop rules.

@Merocle DM me and I'll share a few links if this is a use-case you'd like to explore.

There's probably a secondary market amongst music producers for this. USB MIDI clock signals frequently suffer from jitter due to operating system scheduling and USB buffer, leading to inconsistent timing (often 2–20+ ms variation) when syncing DAWs to external hardware. If you could mitigate this and have stable external clock come in via USB and out via WIFI or ethernet you'll have solved a real problem for a market that's always looking for an excuse to burn money on stuff like this.

This is a bit serious. I really need your feedback. Long story short, I think Ethernet is a relic of the past and should be confined to walls and floors. Modern laptops and many devices are moving away from it in favor of Wi-Fi, which isn’t always convenient and ends up requiring adapters and dongles on desks. So I thought, what if we made a router with USB ports that essentially act as USB-to-Ethernet adapters inside the router? In my personal experience, this would be super cool for certain use cases. For example, at a conference booth, WiFi or Ethernet could be used as an uplink to the router. And the router could have a VPN set up to the company network. For a HomeLab, this is a pretty neat setup as a demo stand. And it would work just fine as a travel router, too. Would you like to see a device like this come to life? This could very well be the first hardware product from developers, made for developers. I mean, from JetBrains. On the technical side, I have some rough ideas for features and unique capabilities. For example, automatic detection of the WAN port regardless of where you plug in the network (the device has 2 Ethernet ports). I’d be happy to dive into the details in the near future. I’d appreciate it if you shared this post - if only because it adds a bit of variety to your feed, which I’m sure is full of AI content and news about AI, offering something tangible and practical.