惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

Y
Y Combinator Blog
D
Docker
有赞技术团队
有赞技术团队
D
DataBreaches.Net
The GitHub Blog
The GitHub Blog
爱范儿
爱范儿
H
Help Net Security
美团技术团队
MyScale Blog
MyScale Blog
B
Blog RSS Feed
C
Check Point Blog
Microsoft Security Blog
Microsoft Security Blog
阮一峰的网络日志
阮一峰的网络日志
A
About on SuperTechFans
小众软件
小众软件
Cyber Security Advisories - MS-ISAC
Cyber Security Advisories - MS-ISAC
GbyAI
GbyAI
G
Google Developers Blog
月光博客
月光博客
Google DeepMind News
Google DeepMind News
H
Hackread – Cybersecurity News, Data Breaches, AI and More
Blog — PlanetScale
Blog — PlanetScale
MongoDB | Blog
MongoDB | Blog
F
Fortinet All Blogs

seangoedecke.com RSS feed

Tell agents the why, not just the how Slow developer experience will bottleneck fast models AI is breaking our proxies for expertise Don't build tools for AI agents They really do think AI might kill everyone Why we should anthropomorphize AI agents Automatically detecting AI text in my browser Radical responsibility means treating people like tools How to protect yourself from workslop You have to beat the models at something Selling out You should never be angry at work Readers can't identify watermarked AI text Good writing is obvious, not original Help peer AI text watermarking is not a big deal No, local models will not win Advanced AI sycophancy I got an email about resistance How to keep thinking Giving and taking credit in big tech companies AI models need moral support to make discoveries You don't have to be smart if you can think clearly LLMs reward expertise Powerful AIs might escape containment by releasing themselves as open-weight models Impro is a handbook for running a cult Overtraining as the path to human-like AI What does "playing politics" mean for software engineers? In defense of not understanding your codebase Blog about things you don't understand yet
LLM-generated skills work, if you generate them afterwards
2026-02-17 · via seangoedecke.com RSS feed

LLM “skills” are a short explanatory prompt for a particular task, typically bundled with helper scripts. A recent paper showed that while skills are useful to LLMs, LLM-authored skills are not. From the abstract:

Self-generated skills provide no benefit on average, showing that models cannot reliably author the procedural knowledge they benefit from consuming

For the moment, I don’t really want to dive into the paper. I just want to note that the way the paper uses LLMs to generate skills is bad, and you shouldn’t do this. Here’s how the paper prompts a LLM to produce skills:

Before attempting to solve this task, please follow these steps: 1. Analyze the task requirements and identify what domain knowledge, APIs, or techniques are needed. 2. Write 1–5 modular skill documents that would help solve this task. Each skill should: focus on a specific tool, library, API, or technique; include installation/setup instructions if applicable; provide code examples and usage patterns; be reusable for similar tasks. 3. Save each skill as a markdown file in the environment/skills/ directory with a descriptive name. 4. Then solve the task using the skills you created as reference

The key idea here is that they’re asking the LLM to produce a skill before it starts on the task. It’s essentially a strange version of the “make a plan first” or “think step by step” prompting strategy. I’m not at all surprised that this doesn’t help, because current reasoning models already think carefully about the task before they begin.

What should you do instead? You should ask the LLM to write up a skill after it’s completed the task. Obviously this isn’t useful for truly one-off tasks. But few tasks are truly one-off. For instance, I’ve recently been playing around with SAEs and trying to clamp features in open-source models, a la Golden Gate Claude. It took a while for Codex to get this right. Here are some things it had to figure out:

  • Extracting features from the final layernorm is too late - you may as well just boost individual logits during sampling
  • You have to extract from about halfway through the model layers to get features that can be usefully clamped
  • Training a SAE on ~10k activations is two OOMs too few to get useful features. You need to train until features account for >50% of variance

Once I was able (with Codex’s help) to clamp an 8B model and force it to obsess about a subject1, I then asked Codex to summarize the process into an agent skill2. That worked great! I was able to spin up a brand-new Codex instance with that skill and immediately get clamping working on a different 8B model. But if I’d asked Codex to write the skill at the start, it would have baked in all of its incorrect assumptions (like extracting from the final layernorm), and the skill wouldn’t have helped at all.

In other words, the purpose of LLM-generated skills is to get it to distil the knowledge it’s gained by iterating on the problem for millions of tokens, not to distil the knowledge it already has from its training data. You can get a LLM to generate skills for you, so long as you do it after the LLM has already solved the problem the hard way.


If you liked this post, consider subscribing to email updates about my new posts, or sharing it on Hacker News.

Here's a preview of a related post that shares tags with this one.

AI interpretability has the same problems as philosophy of mind

If you want to know how a particular AI language model works, the current state-of-the-art approach is to use a “probe”. Pick one layer of the model - a single region of its “brain” - and train a smaller classifier model to see if there are neuron patterns that light up reliably in response to particular concepts. This is the core technique of a field called “mechanistic interpretability”.
Continue reading...