Pinned Tweet
I wrote up how I built the shitty robot so you can too. This was a fun project that will keep on giving.
Thanks to all the open weights folks out there, without whom this would not have been possible.
mariozechner.at/posts/2026-05-…
Thanks to all the open weights folks out there, without whom this would not have been possible.
mariozechner.at/posts/2026-05-…
30
62
791
326,103
had this built by a spare agent. just to one up armin. told it to also do an eval. bunch of tasks different models.should.complete using this, quickjs/js and armin's djs/TS.
it uses codemode do run the eval, spawning subagents in the same pi durable session and i think that's beautiful.
it uses codemode do run the eval, spawning subagents in the same pi durable session and i think that's beautiful.
after short deliberation, i'm now going to built my own language to be used with codemode. i'll call it pico.
x.com/i/broadcasts/2…
7
2
56
10,854
yes, my prompts are bespoke and well informed, thanks for asking. let's hillclimb until ultimate success.
a type checked codemode language does improve correctness, currently at the cost of initial attempt failures. that can be fixed with improvements to the language and the ssed to the agent.
a type checked codemode language does improve correctness, currently at the cost of initial attempt failures. that can be fixed with improvements to the language and the ssed to the agent.
1
0
1
1,106
the goal here is a durable execution vm with smallest possible snapshot sizes, so we can get rid of the replay based durable execution.
0
0
1
911
Mario Zechner retweeted
If I want someone to get something done at Earendil I just nerdsnipe them with my clanker. x.com/badlogicgames/…
had this built by a spare agent. just to one up armin. told it to also do an eval. bunch of tasks different models.should.complete using this, quickjs/js and armin's djs/TS.
it uses codemode do run the eval, spawning subagents in the same pi durable session and i think that's beautiful.
3
3
43
4,460
Mario Zechner retweeted
the slop list works folks. don’t grow your following by being a reply guy and slinging ai slop around.
grow your following by posting authentic and meaningful contributions
grow your following by posting authentic and meaningful contributions
89
21
771
42,185
Mario Zechner retweeted
Are Agents soon going to have only 1 tool? We went from a handful of tools, to dynamic tool search, and now to collapsing everything into code mode.
Wrote down how Code Mode works, why a typed JS/Python sandbox can beat raw bash, and why MCP is going to go back strong in 2027. ple for Gemini 4 Argon.
philschmid.de/code-mode
If you plan to explore code mode i can recommend testing with pi! Easily hackable and comes with strong foundation.
Wrote down how Code Mode works, why a typed JS/Python sandbox can beat raw bash, and why MCP is going to go back strong in 2027. ple for Gemini 4 Argon.
philschmid.de/code-mode
If you plan to explore code mode i can recommend testing with pi! Easily hackable and comes with strong foundation.
56
18
264
15,905
Mario Zechner retweeted
I got a lot of inspiration talking to the brightest minds in the fields @badlogicgames and @mitsuhiko.
At @deepseek_ai we believe in scaling across agent harnesses and aim to provide good support to @pidotdev among all the major harnesses in the world in each generation of our models.
At @deepseek_ai we believe in scaling across agent harnesses and aim to provide good support to @pidotdev among all the major harnesses in the world in each generation of our models.
recommended reading. had the good fortune to talk with @tianyi a little while ago, and the DSH folks are one of the not so many teams doing innovative stuff in the harness space.
the below largely aligns with what we are gravitating towards at earendil as well. single, long lived entity, ad-hoc, freeform, possibly recursive delegation for actual tasks, cross-system event ingestion and triggers.
good stuff.
28
30
593
54,902
Mario Zechner retweeted
@badlogicgamesh@tianyive been building (on top of Pi Durable) is going in this direction a lot... I am just trying to validate the idea (building it for fun) but more and more write ups about similar models are popping out and sort of validating the idea at a high level at least.
It is like this multi-threaded harness where you can use "main" as the control pane that coordinates threads and is always unblocked. All are infinite chats with memory. x.com/giuseppegurgon…
It is like this multi-threaded harness where you can use "main" as the control pane that coordinates threads and is always unblocked. All are infinite chats with memory. x.com/giuseppegurgon…
♾️ infinite chat with memory built-in
⎇ virtually infinite parallel work is done in threads
! main orchestrates threads or you can @:<thread> them
1/3 todos bar in the footer
5
4
63
9,519
after short deliberation, i'm now going to built my own language to be used with codemode. i'll call it pico.
x.com/i/broadcasts/2…
10
4
67
15,110
after short deliberation, i'm now going to built my own language to be used with codemode. i'll call it pico.
x.com/i/broadcasts/2…
x.com/i/broadcasts/2…
23
2
214
37,677
haven't managed to take his tokens away yet. we are at "let's build our own JS interpreter in rust" levels of ai psychosis now.
Opus Induced AI Psychosis x.com/i/broadcasts/2…
9
2
180
15,588
love this.
Weekend hack: HackerTube. I was annoyed with YouTube's recommendations so I had Claude filter down 43M comments over 20 years of HN to 6,454 videos that keep being recommended. *Significantly* better viewing experience, ~2 hours with Opus 5.5:
hackertube.org
- Excellent (state is kept in local browser session)
- Thumbnail is a representative frame from middle of the video instead of the super duper A/B test hyper-optimized user uploads
- Underlying data from the comment analysis as one zip file, point your Claude at it and ask for your own version
- HackerTube Longs ✊
A bit more on the process:
I use Claude Code (desktop app for Mac), Opus 5.5 on Medium effort as my default. I only talk to Claude and have it do everything, the only "manual work" I helped with in tandem was in the browser (reminiscent of MenuGen): 1) I had to create some kind of API key from Google Cloud UI to query YouTube for video metadata, 2) I booted up a VPS box for serving and 3) I bought the domain name. Once the box was up I gave Claude the IP and told it update, secure, harden, deploy and optimize the code, and it did the whole thing.
Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast. What else are we not thinking of.
6
2
172
17,889
recommended reading. had the good fortune to talk with @tianyi a little while ago, and the DSH folks are one of the not so many teams doing innovative stuff in the harness space.
the below largely aligns with what we are gravitating towards at earendil as well. single, long lived entity, ad-hoc, freeform, possibly recursive delegation for actual tasks, cross-system event ingestion and triggers.
good stuff.
the below largely aligns with what we are gravitating towards at earendil as well. single, long lived entity, ad-hoc, freeform, possibly recursive delegation for actual tasks, cross-system event ingestion and triggers.
good stuff.
17
41
768
89,546
sample efficiency is the actual frontier
Yann LeCun and Randall Balestriero just Proved the Platonic Representation Hypothesis, and found a better way to achieve it.
The Platonic Representation Hypothesis suggests that AI models trained on different modalities, like images and text, eventually learn similar world. But this paper shows that how models are trained matters more than simply scaling them up.
So they mathematically show that, under certain conditions, two JEPAs trained completely separately can learn essentially the same underlying representation of the world. One learns from different images of the same scene, while the other learns from different descriptions of that scene, without ever being trained on image-text pairs.
Their approach achieves the same image-text alignment with 42x fewer training samples than next-token prediction. A 109M-parameter text JEPA even outperforms LLMs up to 72B parameters on cross-modal alignment benchmarks.
alphaxiv.org/abs/2610.39302
10
6
150
13,011