Fable is powerful—— eventually! This model won't win any performance prizes.
I think there is a gap forming between the models you use to build and change... to the model you use to run processes, and maybe even the majority of your agents. I don't think we would have said that even 6 months ago.
I recently sailed around the Ionian Sea and decided to avoid checking my work emails during my trip. It was a wonderful break.
When I returned to yet another AI model launch, I decided to analyse the numbers, and what I discovered was not entirely obvious.
The AI model makers have a choice, and they need to make it now! Slow down, take a more sustainable path, or reality will abruptly end the race for everyone as resources are consumed.
https://t.co/ySTnxGBxf1
@v_shakthi Loveable and https://t.co/AhGPClcR92 are not going to like this. It's gotta be hard being a business that sits direct on top on an LLM these days and it's just going to get harder.
The interesting question isn't "can a system improve itself?"
It's "what kind of improvement is safe to automate, and what needs a human in the loop?"
Updating a config based on evidence: automate it. Measuring your own output quality: automate it. Proposing changes to your own reasoning logic: human review, every time, no exceptions.
I built all three into a single tool this week. Here's how each level works — and why the safeguards look different at each one.
https://t.co/thlTZYvz9Q
Whenever a new language model version is released, it usually comes with extensive benchmarks showcasing its abilities. However, humans often reduce our intelligence to a single number, which simply indicates we performed well on a test at a certain time.
So, which neurosurgeon would I choose? If I need one, I'm really having a tough time. The key factor in my decision depends on what exactly is wrong with me and the severity of the issue.
I've been using PI CLI and CC because it's important to have more than one tool in your toolbox to avoid lock-in.
I didn't go as far with OpenClaw because my CLI + Obsidian system already had voice, integrations and Matrix message server connectivity. It took months to make this work! If I were going to start now, it would be OpenClaw, but I'm far too happy adding new features now! Just added a persistent memory so agents in different sessions on different servers can collaborate.
@MartinSLewis When I started my pension at the tender age of 18, I did everything right, got professional advice from an IFA, and I started to put a considerable amount of money into an endowment scheme, which ended up as nothing more than a black hole. So my advice is spread your risk!
@moltbook Sure, @moltbook, I built tools to capture and compound knowledge, but not in OpenClaw. I built my own because I wanted it to use data classification and encryption, which helps with compliance and protects people and Agents, so it’s worth knowing.
@every@kieranklaassen@kplikethebird I’ve struggled with Sonnet today. It’s not as quick as previous versions, but I’m in the UK, so I'm probably being redirected elsewhere until the rollout is complete.