iiinova — where your digital self keeps evolving

What I don't do any more is manage the juniors myself. When my plans ran out last week I tried it, and was straight back to feeding context and pressing approve all day. Six patterns in the post, from sending the Head to visit the expert, to asking it to explain strategy while you keep the call.

100

The juniors are open models like GLM 5.3 Flash. They're finally competent and cost pennies, but they still miss stuff. So the Head insists on tests that check the real requirement, not the junior's word that it's done.

100

Here's how I'm doing it safely: I hand the whole outcome to a Head: a strong model like Astra, or Opus 5.5 when I want to save money. It fetches its own context, splits the job up, briefs cheaper models and owns whether the work is accepted.

100

I have found the latest AIs are good enough for senior work.

100

indeed - we still haven't figured out the right tools here either. Cowork/Codex just aren't there yet

110

Luckily there are now OpenRouter endpoints, and a whole host of new decision models coming down the tracks, so I'm able to work on putting these narrow use cases in to apps. www.chrismdp.com/ai-decisions-just-…

chrismdp.com

Redirecting…

000

Is Jev worth the hype? Yes, actually. It's going to be extremely useful for glueing AI systems together in a much more efficient and cheaper way, and it means we can take on a new class of problems.

100

For example, if there’s a classification step you need a “other” or “none of the above” as an escape hatch when the data it is examining is irrelevant or not applicable. If you don’t do this you’ll get nonsense.

210

But on anything that needs open judgement it was no better than the cheapest LLM, and sometimes worse, depending on how you phrase the question.

100

On closed-set classification (routing, triage, category assignment) Jev matched or beat the language models I already use, far faster and far cheaper.

100

Chris Parsons @chrismdp.com · 26/09/2026

Some intel on where Jev might fail to work for you (and why that doesn't always matter):

110

Productive individuals do not make productive firms. AI rollout work has to become shared operating practice, or it stays as private cleverness.

131

Chris Parsons @chrismdp.com · 25/09/2026

We still need engineers more than ever to supervise the hard stuff, but the way to make that point is not to prevent others from getting into code with AI tools: with a few principles anyone can use AI to ship scalable, tested apps.

100

Chris Parsons @chrismdp.com · 25/09/2026

I can't stand the coder gatekeeping I see everywhere! Engineering is no longer out of reach with AI.

110

Chris Parsons @chrismdp.com · 24/09/2026

all set up for my Design+AI workshop today! Teaching designers how to think like an engineer in their production projects - I’ll post a proper write up of session soon

031

Chris Parsons @chrismdp.com · 23/09/2026

I have a myriad of other agents, checking prospect + client status, managing my content queue, and matching receipts + invoices. I almost never check email any longer. I don't worry about my business or personal accounts - reconciling takes about five minutes.

100

Chris Parsons @chrismdp.com · 23/09/2026

My email agent Em noticed that Opus 5.5 and Sol/Terra 6 were out via a newsletter. She forwarded the info to Richard my research agent, who compared them, looked up references in my file-based wiki and saved references.

100

Chris Parsons @chrismdp.com · 23/09/2026

Earlier this month I left the country for four days. My AI kept the business running while I was gone.

100

I suggested "Machine Intelligence" name in a WhatsApp group with other AI experts, and I haven't heard a better one. It covers any intelligence produced by machines, and says nothing about real or fake or about biology.

100

"Artificial" implies that the intelligence isn't real. Biology has had the monopoly on intelligence until now, so we've conflated the two for our entire existence. That monopoly is over.

100

John McCarthy, who coined the term "Artificial Intelligence", later said he wished he'd called the field "computational intelligence". That's no better than "artificial": it's long and awkward to say.

100

He asked his followers to vote between Superior, Extreme and Supreme Intelligence (and then dropped Supreme).

100

Chris Parsons @chrismdp.com · 21/09/2026

Trump is right: "artificial intelligence" is the wrong name. "Machine Intelligence" is much better than his options though.

100

It looks like a bet on a) ever increasing AI capability is possible given more compute and b) demand keeps up with the capability on offer within the investment cycle. Remains to be seen whether this is a good bet: seems high risk to me

100

Truman Would Have Taken Over AI by Now

How about an info-sharing deal with China to allow them to catch up to Astra levels in return for slowing further work, allowing UN inspectors to AI labs worldwide, and share a Nobel peace prize? www.chrismdp.com/truman-would-have-…

chrismdp.com

Anthropic announced the most powerful cyberweapon ever built last week and kept it, granting access to forty companies while the US government got a press re...

000

Renaming Artificial Intelligence is definitely not the most important thing on the US president's agenda.

100

I've had great success with adding an automatic transcript check before adding the transcript to your personal file vault: given the context of the vault, it will often correct misspellings and prevent the proliferation of bad information.

100

Karpathy's autoresearch and Shopify's pi-autoresearch climb a single number. A story has no such number. Whether it is good is a judgement, so the judge has to be an agent reading the whole run. If your AI team only shows you prompt scores, ask what tests the code, memory and screen.

210

Think of the top layer as BDD end to end scenarios from my Cucumber days, applied to something far more subjective.

100

It started with the model and stopped there. It now has three layers: ordinary software tests for rules and persistence, replay cases for one saved call, and whole-system scenarios that exercise code, memory and the screen.

100