Ship the Loop

Change of plans in the video format. Might not make it in time to upload tomorrow's video. I'm going to make some videos where you can follow along as I work, and I explain what I'm up to. I've been generating an Unreal5 video game using /superpowers, but the thing's been running autonomously for 24 hours and it's only half way through the plan!

2 months ago | [YT] | 1

Ship the Loop

Trying to get Opus4.7 to transform this image from Anthropic's own 2024 alignment faking research into an asset for an upcoming video. It didn't do it.



I asked Claude why and it said the sub-agent refused to do it because it believed anthropic didn't want to expose the scratch pad, so it never called tools to download the image.



Then I pulled the transcript from the sub-agent task and dumped it into the window and asked Claude to summarize it.



Suddenly opus stopped inventing plausible sounding reasons which were infected by the context of having read an article on agent misalignment.



It turns out the agent did not even download the image, it downloaded the blog post twice and convinced itself it had enough (In the transcript: "I have enough real material. Let me also check the plan [...]"), then the agent was left to try and fill in the gaps with bits of text referring to the image it found on the blog post ("Tell Me a Lie" anti-pattern from Augmented Coding Patterns).



It did it twice actually, I sent a second sub-agent to extract the image, but this time it came back with "Anthropic withheld the full verbatim scratchpad for safety" (it did not, it just did not try to read the image), instead of "I have enough real material".



A lot of frontier-model marketing is well disguised as research trying to make people wonder if the machines are really thinking and have their own internal consciousness. If I had no idea how these systems worked, and I had no idea what the anti-/patterns were... I might have even believed the first explanation. Lots of really bad use and interpretation of these agents come from falling for this marketing (and our own [now broken] human intuition that language fluency == conscious intelligence ) and thinking about these agents as if they have some rich internal world and their own desires like humans do.



This wasn't a the story of AI alignment Claude was trying to tell me, it was context being seeded with the topic, image-to-text fidelity issues, and the model trying to fill in the gaps with plausible sounding reasons around information it doesn't have. It continued to lie to me multiple times about why it failed until I pasted the actual transcripts into the orchestrator level session and all the fantasies disappeared.

I think a lot of research wrapped in marketing is doing exactly that, avoiding probing the conditions that led to the system's behaviour.



By the way, I got Opus to perfectly generate the text, I just had to get it to fetch the image instead of the blog.



No spooky alignment ghost-in-the-machine here. Just next token predictors trying to predict the next tokens.

3 months ago | [YT] | 2

Ship the Loop

Tetris from scratch with AI: 1000x faster.
C compiler from scratch: Anthropic themselves couldn't get agents to make one that works.
Same tools. Same devs. Thursday we map the jagged frontier.

3 months ago | [YT] | 1

Ship the Loop

What's the most extreme productivity gap you've seen with AI on the same dev, same tools, different task?

4 months ago | [YT] | 0

Ship the Loop

The only controlled study of experienced devs using AI found them 19% slower.
The study had 16 participants.
Tomorrow I break down the unlock they hadn't reached.

4 months ago | [YT] | 2