@duarte@hachyderm.io

@trentshell @jkb @jneen Even if LLMs, as they get better, create fewer bugs than humans, we’re still bottlenecked by verification and understanding. When writing code, understanding develops naturally, but with LLMs it must be done explicitly. The problem is that their default usage mode doesn’t really encourage this (which is fine for prototyping but not for production). (One can argue that “dark factories” are the way forward, but I’m not sure how well an organization that doesn’t have knowledge about its systems will fare. It will be constrained by whatever the LLM capabilities of summarizing and modifying the system are at any given point in time, beholden to whoever produced the model. And for sure the organization won’t be able to innovate, as LLM aren’t good at extrapolating outside their training data.)

1

0

Duarte @duarte@hachyderm.io

· 44mo ago

Cuties on a branch.

#wildlife #WildLifePhotography #BirdPhotography #photography #Photomonday

hachyderm.io

5

0

2

0

Duarte @duarte@hachyderm.io

· 5mo ago

Replying to

@aphyr Not sure if you've seen it, but this study also corroborates the effects of deskilling: https://arxiv.org/abs/2604.04721

arXiv.org

AI Assistance Reduces Persistence and Hurts Independent Performance

People often optimize for long-term goals in collaboration: A mentor or companion doesn't just answer questions, but also scaffolds learning, tracks progress, and prioritizes the other person's growth over immediate results. In contrast, current AI systems are fundamentally short-sighted collaborators - optimized for providing instant and complete responses, without ever saying no (unless for safety reasons). What are the consequences of this dynamic? Here, through a series of randomized control

0

Duarte @duarte@hachyderm.io

· 5mo ago

Alignment Whack-a-Mole: Finetuning Activates Verbatim Recall of Copyrighted Books in Large Language Models: https://arxiv.org/abs/2603.20957

The paper gets bonus points for featuring Murakami.

arXiv.org

Alignment Whack-a-Mole : Finetuning Activates Verbatim Recall of Copyrighted Books in Large Language Models

Frontier LLM companies have repeatedly assured courts and regulators that their models do not store copies of training data. They further rely on safety alignment strategies via RLHF, system prompts, and output filters to block verbatim regurgitation of copyrighted works, and have cited the efficacy of these measures in their legal defenses against copyright infringement claims. We show that finetuning bypasses these protections: by training models to expand plot summaries into full text, a task

0

Duarte @duarte@hachyderm.io

· 5mo ago

@GossiTheDog@cyberplace.social I guess one part of tokenmaxxing is getting employees to produce traces for training the companies’ models, and the other part is like return-to-office policies, with leadership having a stake in hyperscalers like they had in real estate funds.

0

1

0

Duarte @duarte@hachyderm.io

· 5mo ago

The buzz of the refrigerator pushed against the heavy silence of 3 AM. I stood in the bedroom a few feet away, at the night's most pivotal moment. Belly in, tensing, I focused on slowing both breath and heartbeat. I let past fears and future expectations slip from my perception. I pulled my right arm back, contracting muscles, conjuring strength. It happened in the space of a blink. The wall felt cool against my hand, throbbing slightly with pain, with release. I removed my hand, letting relaxation and peace overtake me as I saw the compressed body of the mosquito, now a red and black stain on the eggshell wall. It was done. I fell back into bed, tired and proud. Afraid, the refrigerator surrendered to silence.

0

Duarte @duarte@hachyderm.io

· 5mo ago

Replying to

@colincornaby@mastodon.social Yep: https://arxiv.org/abs/2603.20957

arXiv.org

Alignment Whack-a-Mole : Finetuning Activates Verbatim Recall of Copyrighted Books in Large Language Models

Frontier LLM companies have repeatedly assured courts and regulators that their models do not store copies of training data. They further rely on safety alignment strategies via RLHF, system prompts, and output filters to block verbatim regurgitation of copyrighted works, and have cited the efficacy of these measures in their legal defenses against copyright infringement claims. We show that finetuning bypasses these protections: by training models to expand plot summaries into full text, a task

0