Dmitro Cashback

AI, agents and emerging tech
sharing what I learn, test and discover
dms open | dyor

[Dmitro Cashback @DmitroCP]

Karpathy has an auto-researcher running on his own machine, and the most useful part of the interview is where he says it stops working

the idea is the one every frontier lab is chasing: LLMs improving LLMs. he calls his own version a playpen for it. one loop, left running with nobody watching, went at a repo he had already tuned by hand and still found improvements

what is actually in it:

→ a single loop, arranged so it can just keep going indefinitely
→ the labs run the same idea across clusters of tens of thousands of GPUs
→ the direction is to experiment on smaller models and take researchers out of the loop entirely
→ the whole thing is steered by a markdown file he wrote describing how the researcher should behave
→ his framing: a research organisation is a set of markdown files describing the roles and how they connect
→ his contest idea: same hardware, different markdown files, see which produces the most improvement, then hand that data back to the model and let it write a better one

three things people quoting this will skip:

the hard limit, in his own words: if you can't evaluate it, you can't auto-research it. rewriting a kernel to run faster with identical behaviour is a perfect fit. most work is not that

the second caveat is about the agent itself. he describes talking to something that is simultaneously a brilliant lifelong systems programmer and a ten year old, and admits he still gets frustrated with agents constantly, especially when one burns compute on a problem it should have spotted

and the honest one: he says the progression is obvious, but you cannot let it fully run yet, because either it genuinely does not work or it is a skill issue nobody has solved. he does not claim to know which

the catch: labs can improve whatever carries a reward. did it compile, did the test pass. anything outside that loop is not being optimised at all, so it stays stuck. that is where the jaggedness comes from

his expected fix is speciation, models shaped for niches instead of one oracle that knows everything. his read on today: we are still in a monoculture

save this one, then read what a long agent run actually costs below ↓

Andrej Karpathy on No Priors

[Dmitro Cashback @DmitroCP]

@0xCodez Why Alex looks so amazing

[Dmitro Cashback @DmitroCP]

Karpathy has an auto-researcher running on his own machine, and the most useful part of the interview is where he says it stops working

the idea is the one every frontier lab is chasing: LLMs improving LLMs. he calls his own version a playpen for it. one loop, left running with nobody watching, went at a repo he had already tuned by hand and still found improvements

what is actually in it:

→ a single loop, arranged so it can just keep going indefinitely
→ the labs run the same idea across clusters of tens of thousands of GPUs
→ the direction is to experiment on smaller models and take researchers out of the loop entirely
→ the whole thing is steered by a markdown file he wrote describing how the researcher should behave
→ his framing: a research organisation is a set of markdown files describing the roles and how they connect
→ his contest idea: same hardware, different markdown files, see which produces the most improvement, then hand that data back to the model and let it write a better one

three things people quoting this will skip:

the hard limit, in his own words: if you can't evaluate it, you can't auto-research it. rewriting a kernel to run faster with identical behaviour is a perfect fit. most work is not that

the second caveat is about the agent itself. he describes talking to something that is simultaneously a brilliant lifelong systems programmer and a ten year old, and admits he still gets frustrated with agents constantly, especially when one burns compute on a problem it should have spotted

and the honest one: he says the progression is obvious, but you cannot let it fully run yet, because either it genuinely does not work or it is a skill issue nobody has solved. he does not claim to know which

the catch: labs can improve whatever carries a reward. did it compile, did the test pass. anything outside that loop is not being optimised at all, so it stays stuck. that is where the jaggedness comes from

his expected fix is speciation, models shaped for niches instead of one oracle that knows everything. his read on today: we are still in a monoculture

save this one, then read what a long agent run actually costs below ↓

Andrej Karpathy on No Priors

[Valentin @martynov014]

https://t.co/dzgA9GUi0l

[Dmitro Cashback @DmitroCP]

@martynov014 Andrey so smart AI guy

[Dmitro Cashback @DmitroCP]

an OpenAI engineer running Codex just said the quiet part: he does far less context engineering than he did a year ago

his slide is three words. context lives outside of the repo. the job stopped being how do i phrase this and became what can it reach

what is actually in the talk:

→ the arc from autocomplete to pair programming to delegation, and why delegation breaks differently
→ plugins as the way to hand an agent context: GitHub, Slack, Notion, Linear, Drive, Calendar, Gmail, SharePoint, Teams, Outlook
→ over a hundred of them now, and most are not coding specific
→ working from inside Slack, where he tags the agent on a real message instead of opening a tool
→ it resolves who a colleague is, finds the doc they posted, and knows which repo it belongs in
→ it applies the team's own style guide so the review loop gets shorter

three things people get wrong about this:

context engineering as a discipline was mostly a workaround for models that could not reach anything. give the agent real access and most of the ceremony disappears

the plugin list is the actual product. a hundred connectors is not a feature page, it is the difference between an agent that drafts and an agent that finishes

and the unglamorous one: he says the highest-impact habit is feeding your review feedback back into the guidelines. almost nobody does it

the catch: every one of those connectors is a permission. the same reach that makes delegation work is the reach you now have to scope, approve and audit

for a year the advice was to write better prompts. the people shipping this are writing better access

save this one, then read what long agent runs actually cost below ↓

guest lecture by Dominik Kundel, OpenAI

[Dmitro Cashback @DmitroCP]

@martynov014 love moment at 39:02

[Dmitro Cashback @DmitroCP]

@AnatoliKopadze So smart man on the video

[Dmitro Cashback @DmitroCP]

@martynov014 Dominik Kundel

his profile @dkundel

[Dmitro Cashback @DmitroCP]

an OpenAI engineer running Codex just said the quiet part: he does far less context engineering than he did a year ago

his slide is three words. context lives outside of the repo. the job stopped being how do i phrase this and became what can it reach

what is actually in the talk:

→ the arc from autocomplete to pair programming to delegation, and why delegation breaks differently
→ plugins as the way to hand an agent context: GitHub, Slack, Notion, Linear, Drive, Calendar, Gmail, SharePoint, Teams, Outlook
→ over a hundred of them now, and most are not coding specific
→ working from inside Slack, where he tags the agent on a real message instead of opening a tool
→ it resolves who a colleague is, finds the doc they posted, and knows which repo it belongs in
→ it applies the team's own style guide so the review loop gets shorter

three things people get wrong about this:

context engineering as a discipline was mostly a workaround for models that could not reach anything. give the agent real access and most of the ceremony disappears

the plugin list is the actual product. a hundred connectors is not a feature page, it is the difference between an agent that drafts and an agent that finishes

and the unglamorous one: he says the highest-impact habit is feeding your review feedback back into the guidelines. almost nobody does it

the catch: every one of those connectors is a permission. the same reach that makes delegation work is the reach you now have to scope, approve and audit

for a year the advice was to write better prompts. the people shipping this are writing better access

save this one, then read what long agent runs actually cost below ↓

guest lecture by Dominik Kundel, OpenAI

[Valentin @martynov014]

https://t.co/dzgA9GUi0l

[Dmitro Cashback @DmitroCP]

@martynov014 saved this buddy

[Dmitro Cashback @DmitroCP]

Y Combinator put a room of frontier researchers together to argue about the least glamorous layer in AI, and the number they opened with should end the argument

harnesses get called scaffolding. wrappers. prompt engineering with extra steps. not real research

then someone runs the same model weights through two different harnesses and the score moves 65 points

what is actually in the session:

→ 04:27 - an auto-researcher that got built by accident
→ 13:56 - harnesses that rewrite themselves
→ 18:35 - Prime Agent, a self-improving RLM harness
→ 21:50 - treating context as an L1, L2, L3 cache
→ 28:33 - how agents actually message each other
→ 30:04 - the ARC-AGI results
→ 43:53 - a local stack running 800x cheaper than the cloud

three things people miss when they skim this:

the 30 to 95 jump is on the private holdout, the set nobody outside the organisers can train against. that is the version of the benchmark that still means something, and the harness is what moved it

it is not one magic wrapper either. going from harness one to harness two moved results by 18% on its own. the wrapper has versions, and the spread between them beats most model upgrades

and one team's entry reached 100%. same benchmark, no new model

the catch: every number here is a harness result on a benchmark, and a benchmark hands you the task already defined. none of it tells you what happens when the job is vague, which is most jobs

but the direction is hard to argue with. everyone has the same engine now. what you bolt around it is the part still moving

save this one, it is an hour and it is free

session: Y Combinator Paper Club, "Why The Harness Matters More Than The Model"

[Dmitro Cashback @DmitroCP]

want to hear it irl

[Dmitro Cashback @DmitroCP]

Y Combinator put a room of frontier researchers together to argue about the least glamorous layer in AI, and the number they opened with should end the argument

harnesses get called scaffolding. wrappers. prompt engineering with extra steps. not real research

then someone runs the same model weights through two different harnesses and the score moves 65 points

what is actually in the session:

→ 04:27 - an auto-researcher that got built by accident
→ 13:56 - harnesses that rewrite themselves
→ 18:35 - Prime Agent, a self-improving RLM harness
→ 21:50 - treating context as an L1, L2, L3 cache
→ 28:33 - how agents actually message each other
→ 30:04 - the ARC-AGI results
→ 43:53 - a local stack running 800x cheaper than the cloud

three things people miss when they skim this:

the 30 to 95 jump is on the private holdout, the set nobody outside the organisers can train against. that is the version of the benchmark that still means something, and the harness is what moved it

it is not one magic wrapper either. going from harness one to harness two moved results by 18% on its own. the wrapper has versions, and the spread between them beats most model upgrades

and one team's entry reached 100%. same benchmark, no new model

the catch: every number here is a harness result on a benchmark, and a benchmark hands you the task already defined. none of it tells you what happens when the job is vague, which is most jobs

but the direction is hard to argue with. everyone has the same engine now. what you bolt around it is the part still moving

save this one, it is an hour and it is free

session: Y Combinator Paper Club, "Why The Harness Matters More Than The Model"

[Dmitro Cashback @DmitroCP]

Anthropic engineer stood on a conference stage and did the thing almost no vendor does: she put a number on what autonomy costs

the mode she was describing is one flag. you let Claude execute instead of asking, and the permission prompts stop. that single change is what turns a chat window into something that runs a loop, a team of agents, or a job overnight

what is actually behind that flag:

→ no permission prompts, the model executes instead of proposing
→ it still refuses destructive actions and anything posted externally without asking
→ underneath sits a screening layer of models and classifiers checking an action for danger before allowing or blocking it
→ it reads back through your own prompts to see whether you actually asked for the risky thing
→ this is the piece that makes overnight runs and agent teams work at all

three things the clips going around leave out:

the price. she said it on stage: 10 to 40 percent more depending on the model. her own slide is blunter and calls it a different cost class than interactive use

why it costs that. an unblocked agent does more of everything. more reads, more retries, more verification, more subagents. nobody charges you for the permission prompt, they charge you for what happens once you remove it

when it bites. this flag is what gets you from one session to twenty or thirty. the bill arrives at exactly the moment the setup starts being useful

the catch: this is not a safety toggle, it is a budget decision wearing a safety toggle's clothes. the guardrails stay on. what changes is how much work happens between your approvals, and you are paying for all of it

her framing for the year is the part worth keeping. 2025 was about getting information into the model. 2026 is about getting it back out, to you

save this before you turn the flag on ↓

clip from Daisy Hollman's talk at NDC Copenhagen

[Dmitro Cashback @DmitroCP]

Anthropic engineer stood on a conference stage and did the thing almost no vendor does: she put a number on what autonomy costs

the mode she was describing is one flag. you let Claude execute instead of asking, and the permission prompts stop. that single change is what turns a chat window into something that runs a loop, a team of agents, or a job overnight

what is actually behind that flag:

→ no permission prompts, the model executes instead of proposing
→ it still refuses destructive actions and anything posted externally without asking
→ underneath sits a screening layer of models and classifiers checking an action for danger before allowing or blocking it
→ it reads back through your own prompts to see whether you actually asked for the risky thing
→ this is the piece that makes overnight runs and agent teams work at all

three things the clips going around leave out:

the price. she said it on stage: 10 to 40 percent more depending on the model. her own slide is blunter and calls it a different cost class than interactive use

why it costs that. an unblocked agent does more of everything. more reads, more retries, more verification, more subagents. nobody charges you for the permission prompt, they charge you for what happens once you remove it

when it bites. this flag is what gets you from one session to twenty or thirty. the bill arrives at exactly the moment the setup starts being useful

the catch: this is not a safety toggle, it is a budget decision wearing a safety toggle's clothes. the guardrails stay on. what changes is how much work happens between your approvals, and you are paying for all of it

her framing for the year is the part worth keeping. 2025 was about getting information into the model. 2026 is about getting it back out, to you

save this before you turn the flag on ↓

clip from Daisy Hollman's talk at NDC Copenhagen

[Codez @0xCodez]

https://t.co/eJB3TgyJwV

[Dmitro Cashback @DmitroCP]

@martynov014 bookmarked this genius woman

[Dmitro Cashback @DmitroCP]

someone just shipped a tool that rewrites Super Nintendo games while you are still playing them

not a rom hack you patch and reload. you type one sentence, the agent walks into the running game, finds the memory structures itself, writes the patch, tests it, and hands the level back to you mid-session

it is called SuperAstra. built on BizHawk, driven by GPT-6 Astra, source is public

what is actually in it:

→ you describe the change in plain english. no assembly, no hex editor
→ the agent inspects the live game and locates the memory it needs on its own
→ it writes the routine or a cartridge patch, then tests its own work before handing it back
→ it remembers what it learned for that exact cartridge version
→ your original ROM file is never modified
→ Windows or Linux, python 3.10, BizHawk, a ROM, an API key

one prompt in the demo: replace level one with a short obstacle course, good number of coins and koopas

what came back was validated at 69 coins, 12 red and green koopas, five screens, a goal gate at the end

three things the clips going around leave out:

look at the status bar in his own demo. that single level swap burned 1,173,137 input tokens and returned 6,501. for one sentence

someone in his replies is rebuilding every character in Street Fighter II Turbo and is almost a billion tokens deep

and it loses. he typed kill lavos and win the fight, watched chrono die over and over, gave up, and asked it to skip to the ending instead. his own words: lavos was too powerful for this to work

the catch: this is the best reverse engineering demo anyone has posted this week, and it is still a model guessing at raw machine code with a meter running. one reply under it just says every change made the game worse

the shift here is not modding. it is that reading a binary nobody ever documented stopped being a specialist skill

save this one, the repo is public

[Dmitro Cashback @DmitroCP]

@grokkedd Sam Altman real smart...

[Dmitro Cashback @DmitroCP]

@martynov014 this token budget looks good

[Dmitro Cashback @DmitroCP]

@martynov014 hah yeeep

[Dmitro Cashback @DmitroCP]

someone just shipped a tool that rewrites Super Nintendo games while you are still playing them

not a rom hack you patch and reload. you type one sentence, the agent walks into the running game, finds the memory structures itself, writes the patch, tests it, and hands the level back to you mid-session

it is called SuperAstra. built on BizHawk, driven by GPT-6 Astra, source is public

what is actually in it:

→ you describe the change in plain english. no assembly, no hex editor
→ the agent inspects the live game and locates the memory it needs on its own
→ it writes the routine or a cartridge patch, then tests its own work before handing it back
→ it remembers what it learned for that exact cartridge version
→ your original ROM file is never modified
→ Windows or Linux, python 3.10, BizHawk, a ROM, an API key

one prompt in the demo: replace level one with a short obstacle course, good number of coins and koopas

what came back was validated at 69 coins, 12 red and green koopas, five screens, a goal gate at the end

three things the clips going around leave out:

look at the status bar in his own demo. that single level swap burned 1,173,137 input tokens and returned 6,501. for one sentence

someone in his replies is rebuilding every character in Street Fighter II Turbo and is almost a billion tokens deep

and it loses. he typed kill lavos and win the fight, watched chrono die over and over, gave up, and asked it to skip to the ending instead. his own words: lavos was too powerful for this to work

the catch: this is the best reverse engineering demo anyone has posted this week, and it is still a model guessing at raw machine code with a meter running. one reply under it just says every change made the game worse

the shift here is not modding. it is that reading a binary nobody ever documented stopped being a specialist skill

save this one, the repo is public