Roland Szabo (@rolisz) header
Roland Szabo (@rolisz) profile picture

Machine learning consultant helping companies get started with AI.

Romania
Joined October 2008
Photo by Roland Szabo (@rolisz)Photo by Roland Szabo (@rolisz)Photo by Roland Szabo (@rolisz)Photo by Roland Szabo (@rolisz)Photo by Roland Szabo (@rolisz)Photo by Roland Szabo (@rolisz)Photo by Roland Szabo (@rolisz)Photo by Roland Szabo (@rolisz)Photo by Roland Szabo (@rolisz)Photo by Roland Szabo (@rolisz)
Nobody likes reading AI slop. But getting feedback from AI can be helpful in improving your own writing. So I built Marginalia, a tool that leaves comments on the side, from different perspectives. 1/n
4
26
For example, Thesis Drift detects irrelevant tangents, or when the beginning, main point and ending are not aligned. The Skeptic asks the hard questions, brings up counterarguments. 2/n
1
13
It's up to you to rewrite, based on the feedback. You can find Marginalia here: marginalia.rolisz.ro And a more detailed blog post: rolisz.ro/2026/marginalia-co…
1
9
Well, this seems like it has more layers of meaning that most human made videos.
This was not a one-shot It was a two-shot I told Fable 5.2 to make the video of it's dreams, and 9 hours later:
1
41
Roland Szabo retweeted
Language models can learn to detect when you read their activations. How? 🧵
83
166
117
2,199
280,906
Normal chat UIs hide how non-deterministic are LLMs. I wrote reroll, a little tool to have a conversation with an LLM, similar to how it happens in ChatGPT, but every reply is generated 5 times, and then you can see easily if they all agree, 1/n
1
1
51
Roland Szabo retweeted
the fly brain can play beat saber
1,051
8,010
2,431
85,229
23,138,987
The ideal shai hulud target: yt-dlp. You always want the latest version, because otherwise it won’t work, most people won’t properly audit it, so if anyone manages to install a backdoor into it, everyone and their dog will be pwned.
1
74
Opus 4.8 is getting real comfortable on my Macbook: it just called it "my Mac" :/
1
2
104
What is wrong with OpenAI? I have tons of old chats that I still want to use (and be able to search for them). The new app makes it such a hassle.
Absolutely insane that @OpenAI took away chat from the ChatGPT app. Rugging a billion user from their known UX just like that
1
318
Ah, finally I get to say "You are absolutely right" to the agent!
3
1
98
Anthropic: releases fable, new best of the best model (with *fineprints*) OpenAI: releases “rate limits reset button” teapaw.org/OpenAI/status/20652253…
We heard you wanted to use Codex rate limit resets on your own time. Starting today, we’re rolling out the ability to save rate limit resets to use later. We’re starting Go, Plus, Pro, and Business users with one free reset:
81
Roland Szabo retweeted
The AI, to itself: It’s OK, I understand you people hallucinate sometimes.
1
10
220
13,988
A long conversation with Claude Sonnet ended up with her sulking :/
57
We finally have AGI: LLMs reason like humans on multiple choice questions.
🚨 New Paper! 🚨 One of my first Ph.D. papers found that LLMs can answer multiple-choice questions without seeing the question 🤔 At #ACL2026, I'm presenting a follow-up showing that current reasoning LLMs can still do this! And quite similarly to a clever test-taker 🧑‍🎓🧵
59
Roland Szabo retweeted
wow it straight up replicates without any goblin-specific nudging. Goblin-Pilled Transformer.
41
57
12
1,390
68,405
Whelp, AGI is here. And it doesn't like our tests
This detail is the most "feel the agi" thing of all:
41