Chris Parsons

AI skills, strategy and products for leaders and organisations

[Chris Parsons @chrismdp]

AI Has A People Problem 1/8: Angela The Addict:

She can't stop starting another AI session. She has 3, then 5, then 15 running at once. She is overwhelmed but feels productive. She is working 7 days a week. She is burning out.

I've been this person before. Some mornings I've felt revulsion towards opening up Claude and talking to AI.

If your most prolific person is working with AI at 2am, don't reward them with more work. Check how they're doing instead and help them find ways to regain balance.

https://t.co/TDpK76QEyL

[Chris Parsons @chrismdp]

Fable 5.2 is out, but after a thorough set of tests, it’s not worth the extra cost over Opus 5.5.

My system ran across a number of completed tasks where Opus 5.5 struggled, and Fable only improved over Opus in one of them. Cost was about 20-30% more than Opus (higher price, fewer tokens). So not worth it for a daily driver over Opus.

[Chris Parsons @chrismdp]

The cheapest project is no project. Just because AI can build it fast doesn’t mean you actually need to build it.

https://t.co/FA3XhF680s

[Chris Parsons @chrismdp]

A successful AI pilot tells me the tech works. That’s like 5% of the story, and never the important bit.

The pilot had enthusiastic people, room to improvise and someone senior watching. But the rest of the organisation has existing work, managers with limited time and colleagues who weren't listening or distracted, etc etc.

Before expanding a pilot, draw out the work that will change. Fund the help people will need. Decide which manager owns the change when the pilot team goes back to its day job.

I wrote AWML to help with this.

https://t.co/f5xHeH2hTR

[Chris Parsons @chrismdp]

I gave AI access to my whole world, but I got super lazy about pointless work or work I should have been doing.

The more a system knows about you, the easier it is to ask it what to do next. It will usually give you a plausible answer. Plausible isn't the same as a good decision or something I don’t actually need.

I now give out small bounded jobs and have a clear set of guides to the agent about what needs me.

https://t.co/tvsKyGkuXR

[Chris Parsons @chrismdp]

@emollick Nice. AI is such a good unlocker of things that previously weren't possible or economic to attempt. If we don't raise our ambition we'll never really get the benefits of it!

[Chris Parsons @chrismdp]

I spoke at Øredev this week about the people you meet when AI rolls out in a team.

It's tempting to label them, laugh at aberrant behaviour and move on. But it's important to remember that each person is protecting something, which is why they act the way they do.

The leadership work starts with finding out what that is (hint: it's mostly about listening). Then you can decide whether they need time, a safer tool, an honest answer about jobs, or a better deal from the gains AI brings.

I'll be writing about each of these personas in the coming days. If you can't wait that long, I've put the slides and a write-up from my speaker notes on my site.

https://t.co/TDpK76QEyL

[Chris Parsons @chrismdp]

When I'm not blogging or keeping up with AI, I work with companies to help them use it effectively. Here's what we did at Instem:

Instem's engineers needed more than another AI demo. We ran focused sessions with the engineering team, then weekly coaching with a smaller group on their own codebase and delivery work. That gave people the chance to turn individual experiments into shared ways of working.

The result included two internal adoption pods and engineers building their own skills and patterns. This made a huge difference to their work.

If you want your business to get the most out of AI, I can help with that!

https://t.co/VsVuDl9tgI

[Chris Parsons @chrismdp]

At Øredev today in Malmö - speaking about how to lead people through the messy maze that is an AI rollout.

First time here and impressed so far. Already had some brilliant conversations with old friends and new. Pictured here with the lovely Hanna from the organising team who is looking after us so well!

[Chris Parsons @chrismdp]

AI can make a team busier and still leave people feeling lost.

When AI handles more of the doing, are your people still learning, and are they reimagining the work? If every handoff and approval remains the same, people are still "doing" but the boring bits only, and work just arrives faster and builds up queues.

Our whole identity was built on doing. Give people room to exercise their judgement and rethink what work they want to do.

Something to try: pick one piece of work. Decide which decisions still need a person and which handoffs can go. Make the questions the job: should we build this at all? Is it worth doing, even when it's easy? You can use my new AWML thing for this process (see links)

https://t.co/gsXWgXyHh8

[John Resig @jeresig]

Seeing this makes the end-state clearer to me: code that can't be read by humans, run on machines that humans can't access, via inaccessible APIs, trained on data that humans didn't create.

LLMs are starting to get really good at manipulating mostly-opaque-to-humans code (see: their improved ability to reverse engineer machine code/assembly). Additionally, folks are accessing more-and-more of their LLMs in cloud-based environments without ever actually interacting with an LLM API (myself included).

I think it's going to become increasingly likely that the end state is one in which you ask an agentic company to produce some result and it does it via [some system completely opaque to humans] and produces [some output in a format that humans are unable to interpret].

It makes me wonder if our current state of competing LLMs cross-training on each other will eventually become literally impossible (not just logistically so) when they become fully self-recursive and are producing output that even other LLMs can't understand, let alone humans.

I think we're in something of a golden period here where humans are still capable of understanding how things work and how things are run (shell commands! python! node.js!) but it's going to quickly become irrelevant. We're already past the previous high-point where we were able to see reasoning traces for the top-level LLMs.

I'm skeptical that open models will ever get to this point as the economics are working against them (I suspect that they will always need human-generated training materials in one form or another). But it does make me wonder if the upper-echelon companies will end up converging into this high-level "black box that produces outputs" singularity of sorts.

This has been hitting home for me lately as I've been using Claude Projects in a cloud-based environment where I can't see transcripts, look at the machine it's running on, nor even know what commands it's running. I don't even know what languages its writing its code in. We're just a couple steps away from code being completely irrelevant.

[Chris Parsons @chrismdp]

@IALatino @emollick What about a bigger model to itself to make it better?

[Chris Parsons @chrismdp]

@emollick Yeah - imagine how this could be used for recursive self improvement - not sure I like where this is heading

[Chris Parsons @chrismdp]

@jburnmurdoch count me as violently agreeing :)

[Chris Parsons @chrismdp]

@omarsar0 @karpathy love this! New interfaces on great content is a super good use of AI

[Chris Parsons @chrismdp]

@levelsio Deep domain knowledge across multiple areas. We are limited only by our taste, ambition and imagination

[Chris Parsons @chrismdp]

@GreyAlien huge congratulations Jake!

[Chris Parsons @chrismdp]

@Steve_Yegge would be very interested to hear what you think of this feedback (and the modelling approach!)

[Chris Parsons @chrismdp]

I used AWML on Steve Yegge's excellent Counting House playbook, which allowed the AI to find and fix a few issues.

Yegge designed the Counting House to put cheap AI models on customer service one kind of case at a time. Each model is scored against people and earns its autonomy level by level. It's a good design, and a useful pattern for lots of types of industrial scale business processes like it, so I drew it in AWML to stress test both the design and the language.

These two findings particularly stood out:

Getting only as far as shadowing humans was about three quarters of the complexity. And demotion gets one sentence. Promotion gets five departments signing off. Going back down needs a rule and a workflow too.

If you use AWML on your own workflows, let me know! I'd love feedback.

https://t.co/O3hRW25n3n

[Chris Parsons @chrismdp]

"What else are we not thinking of?"

^^ THIS is the only limit right now - our imagination and our ambition.

[Andrej Karpathy @karpathy]

Weekend hack: HackerTube. I was annoyed with YouTube's recommendations so I had Claude filter down 43M comments over 20 years of HN to 6,454 videos that keep being recommended. *Significantly* better viewing experience, ~2 hours with Opus 5.5:
https://t.co/JIhXN7OSh1

- Excellent recs without any logins (state is kept in local browser session)
- Thumbnail is a representative frame from middle of the video instead of the super duper A/B test hyper-optimized user uploads
- Underlying data from the comment analysis as one zip file, point your Claude at it and ask for your own version
- HackerTube Longs ✊

A bit more on the process:
I use Claude Code (desktop app for Mac), Opus 5.5 on Medium effort as my default. I only talk to Claude and have it do everything, the only "manual work" I helped with in tandem was in the browser (reminiscent of MenuGen): 1) I had to create some kind of API key from Google Cloud UI to query YouTube for video metadata, 2) I booted up a VPS box for serving and 3) I bought the domain name. Once the box was up I gave Claude the IP and told it update, secure, harden, deploy and optimize the code, and it did the whole thing.

Many awesome projects that would have normally taken weeks/months even can now be few-shotted for breakfast. What else are we not thinking of.