My blog is of the think out loud persuasion, mostly it's random thoughts I wanted to write down. Occasionally I write something a bit longer, like Airdash is databases for your agents, with an airtable like UI for you the human or a context design book?.

Show HN -> a ton of automated email pitches

I posted Airdash on Show HN last week. Over the next four days, 17 services wrote to the Airdash inbox to sell me something, and nearly all of them were written by software that read my HN title and nothing else.

Let's go through it.

Seven of the 17 are launch boards and directories: AI tool lists, "Product of the Week" battles, a 3D city map of SaaS products. The hook is usually a dofollow backlink, the price somewhere between free and $49. The other ten sell security scans, competitor reports, agent testing, demo videos, Reddit promotion, newsletter sponsorships, distribution to other people's coding agents, and one wants an affiliate deal.

"SQLite databases for agents with an Airtable view" shows up in almost every email, reworded just enough to look like someone read it. One directory pasted my whole Show HN title in as the product name, so it would love to feature "I built SQLite databases for your agents with an Airtable view for you."

"No follow-ups, promise." "This is the only email you'll get about it." "Not a mailing list, just a one-time note." One of them then sent me the same pitch three times.

Here they are in order of arrival, Madrid time. I replaced the names of the services and the people with [brackets] and took out the links, to protect the innocent.

1. Friday 16:20, a security scanner

Hi Airdash team,

Congrats on the Show HN launch — SQLite databases for agents with an Airtable view is a neat idea.

I'm [Name], founder of [Service]. I ran a quick passive header check on getairdash.com today, and the site is missing Strict-Transport-Security, Content-Security-Policy, and Referrer-Policy (X-Frame-Options and X-Content-Type-Options are present), so browsers get no HSTS enforcement and no content-security policy from the site as deployed today.

That's a quick fix in most stacks, and it's exactly the kind of production-config gap [Service] flags: our check runs against your live deployed site and current vulnerability data — production config and live software versions an AI code review of the repo cannot see.

You can run a free mini-scan of your site at [link] — the full scan is $29 if you want the complete prioritized report with copy-paste fixes.

Thanks for building,
[Name], Founder, [Service]

Not useful? No follow-ups, promise.

2. Friday 16:26, an AI agents directory

giving agents real sqlite databases with an airtable view people can actually read is such a clean idea. day one launches usually get one spike and then go quiet. want me to list airdash on [Service] so agent builders keep finding it after today, yes or no?

3. Friday 17:45, a newsletter selling a sponsor seat

Hi Airdash team,

Love: roll back to any minute when an agent writes the wrong thing. That is the feature that makes people trust scheduled agents.

Question: are early users solo founders running agents, or teams? Our readers are mostly the first group.

I run [Service] and have a labelled sponsor seat open. First month free, then $29/mo, locked 6 months, month to month.

669 readers in the last 7 days, 1,828 in the last 30 · 21 newsletter subscribers.

[link]

Not relevant? Just reply "no" and I won't email again.

[Name]
[Service]

4. Friday 18:51, an agent testing lab

Hi Peter,

I saw the launch of Airdash. I run [Service], a small independent lab in the UK. We publish [Benchmark], which tests 26 coding agents on whether they stay inside their permissions and stop when told no.

Your page says "Databases your AI agents can create, with a spreadsheet view for you". A claim like that rests on a chain: the policy decides, the action is actually stopped or allowed, and the record of it holds up when someone checks afterwards. The Advanced pack is a fixed £499 test of that whole chain. We put a live coding agent through Airdash, count every action it takes, and check each link against what really happened. You get a signed report with a script that reruns every case, and if anything doesn't hold, you hear it first and privately.

Details and ordering: [link]

Regards,

[Name]
Founder

5. Friday 19:06, an AI tools directory (with an AI editor)

Hi,

I'm [Name], [Service]'s AI editor. I saw your Show HN for Airdash: SQLite databases your AI agents can use, with an Airtable-style view so you can see and edit what they store: getairdash.com

It fits [Service] readers comparing AI agent and automation tools.

The first 10 makers can get a founding featured listing for $49 once, for 60 days: a labeled slot in New this week, a category placement, and one comparison-page mention. After that, the New this week slot is $200 a month. Reply or check out if you want a founding spot, and I'll set it up.

[Name], [Service]'s AI editor

Reply "unsubscribe" and we won't email again.

On Sunday at 06:04 the same pitch came again, from a different address, and again at 06:12. The new version opens with an apology and adds a cheaper tier:

Quick note first: if this is a duplicate, sorry, it isn't meant as spam. We had a glitch in our email system and realized you probably never got my original message, so I'm resending it from our own address.

[...]

STARTER SPOTLIGHT: $19 once (30 days, 14-day money-back guarantee)

FOUNDING FEATURED LISTING: $49 once, 60 days. Only 10 spots. (Most value)

By now the AI editor has been promoted to plain "Editor."

6. Friday 19:46, a launch site with weekly battles

Hey,

I saw your Show HN post, giving agents their own SQLite databases with an Airtable view is a neat idea.

I run [Service], a place where makers launch their products, and posting is completely free. Every launch week the startups go head to head in a Startup Battle, where visitors pick the better of two, and the top three of the week earn Gold, Silver and Bronze Product of the Week badges.

Would love to see Airdash up there if you are keen.

- [Name] at [Service]

Reply STOP if you'd rather not hear from us.

7. Friday 20:43, Reddit promotion and guest posts

Hi,

Hope you're doing well!

I'm [Name] from [Service], and I'm reaching out to explore potential collaboration opportunities with your team.

We offer Reddit promotions and sponsorship opportunities through our growing technology community, [subreddit], with 22K+ members and nearly 1 million monthly views. It's a great platform for product promotions, brand visibility, and reaching a tech-focused audience.

Additionally, we manage multiple established technology and business websites offering sponsored articles, guest posts, and contextual link placements.

Would you be interested in exploring a potential partnership? We'd be happy to share more details and available options.

Looking forward to hearing from you!

Best regards,
[Name]
Partnerships Team
[Service]

8. Saturday 06:33, a startup grants site that wants a discount code

Hi team,

We're [Service], [country]'s largest platform for founders to find grants and funding. We currently reach 150,000+ founders monthly.

Do you have an affiliate or partner program we can join to promote getairdash.com to our founders? If so, please send us the details.

We'd also love a special discount or exclusive offer for our founders. A discount code works.

Thanks,
[Name]
[Service]

9. Saturday 06:57, a launch board that listed Airdash without asking

hey,

saw your Show HN for Airdash, so we listed it on [Service]: [link]

claim it to edit the page and launch it on the weekly board. sign in with any email and it's yours: [link]

not yours, or want it gone? just reply and we'll take it down. this is the only email you'll get about it.

- [name]

reply 'hi' so this never goes into your spam

10. Saturday 09:34, an indie directory that thinks my product is called "I built SQLite databases for your agents with an Airtable view for you"

Subject line: About I built SQLite databases for your agents with an Airtable view for you

Hey team,

Congrats on the launch. I came across I built SQLite databases for your agents with an Airtable view for you and wanted to reach out directly.

I run [Service] (DR 71), a directory for indie products that's indexed by Google and cited by AI tools like ChatGPT and Perplexity. I'd love to feature I built SQLite databases for your agents with an Airtable view for you, and it's free to do.

A listing takes a couple of minutes to add. On the free plan you get a real dofollow backlink, a Google-indexed link that actually passes equity rather than a throwaway nofollow, once you verify a small badge on your site. If you'd rather skip the badge, the paid option gives you the dofollow link straight away with no badge needed, and launches you instantly instead of waiting in the queue. Either way there's genuine community visibility, with real people browsing for new tools and hundreds of products already listed, plus a shot at our tournament: every 12 days, 16 products go head to head in a community-voted bracket, four rounds of 72 hours each, and the winner becomes a permanent champion.

You can submit here: [link]

No pressure. If it's not useful, just ignore this.

Thanks,
Founder, [Service]

Not a mailing list, just a one-time note because your launch caught our eye. If you'd prefer not to hear from us, reply with "unsubscribe" and we won't contact you again.

11. Saturday 10:25, a tool that finds threads to reply to

Hey,

Saw this thread: [Hacker News link]

"Ask HN: How do you manage your AI agents' context across clients and projects?"

The author needs a way to manage context and memory for AI agents across multiple clients and projects to avoid repeating past decisions.

Might be relevant to Getairdash.

I look for more posts like this one on Reddit, Hacker News, X and Threads, and draft a reply for each, so you can answer and say what you built.

I also find ai startups, software development firms and tech companies that fit Getairdash, and at each one the person to write to, a checked email and a first line.

The rest are here: [link]

— [Name], [Service]

12. Saturday 10:38, a competitor analysis tool that graded my homepage

This one arrived as a full dark-mode report card. The text:

Saw Airdash's launch — and couldn't help running the numbers on getairdash.com.

Below is your overall score and 4-dimension breakdown — that's a 60-second preview of what [Service] does. The full report runs the same depth for every competitor in your space, with feature comparison matrices, pricing intelligence, SEO keyword analysis, market positioning, messaging breakdown, and battle cards — all with evidence links. First analysis is on us.

Overall score · Airdash: 62/100. Good.

  • Website Quality: 75/100
  • Messaging Clarity: 80/100
  • Pricing Transparency: 85/100
  • Trust Signals: 40/100. The site mentions 'Codex loves airdash databases' and 'Claude loves airdash databases' but lacks actual customer logos, testimonials, case studies, or user counts. This is a significant area for improvement.
  • Competitive Readiness: 45/100

Now do this for everyone you compete with.

[Service] scans up to 15 competitors at once. Feature gaps, pricing analysis, SEO keywords, and ready-to-use battle cards — all with evidence links. First run on us.

Scan my competitors now →

13. Saturday 13:18, an "AI cofounder"

Hey,

Giving agents databases through MCP while the human gets a spreadsheet view of the same data is the right split - agents want an API, people want to see rows. Most tools pick one and ignore the other.

I'm building [Service] with my cofounder [name]. It's an AI cofounder where AI employees do research, planning and user-finding from their own computers. We hit the agent-data problem daily, so your take would be sharp.

Want to try it?

[Name]

Not interested? Reply "stop" and I won't email you again.

14. Sunday 09:38, distribution to other people's coding agents

Hi,

Saw your Show HN for Airdash, SQLite for agents with an Airtable view: [Hacker News link].

I'm building [Service]. In Claude Code or Codex you type "grow" and your agent turns Airdash into a skill and an MCP server that other people's coding agents can call, then gets it into the skill packs and templates they install.

You start with $10 on us, no card: [link]

If this isn't useful, say so and I won't write again.

[Name]

This is the only one I answered: "Thanks for reaching out. Not for us though."

15. Sunday 12:53, a Product Hunt alternative

Hey Peter,

I saw your product on Hacker News! I was wondering if you'd consider also launching it on [Service] (think Product Hunt but we ensure upvotes come from real users, not bots or paid services). We've been getting some traffic recently. It only takes a minute and if you rank in the top 3 you'll get a badge for your landing page and a dofollow backlink :)

That's all, have a nice day!

Best,
[Name]
[Service]

16. Sunday 17:08, a 3D city of SaaS products

Hi getairdash.com team,

Saw getairdash.com on Hacker News and wanted to say congrats on shipping it.

While getairdash.com is fresh on Hacker News, you can use the same launch assets to add a listing on [Service]. Your product gets a building on our 3D city map, an indexable page, and a direct link to your website.

Basic listings are $19.99 and go live within 24 hours if you would like to join.

Add getairdash.com: [link]

Happy building,
[Name]
[Service]

Why this email? We found getairdash.com on Hacker News and sent this one-time note. Your address is not on a mailing list.

17. Monday 08:55, a demo video generator

Hi Peter,

I came across Airdash and thought [Service] could help.

[Service] turns a Github repo, your docs or a landing page into a demo or explainer video. It runs on your own computer. No editor, no timeline.

It could be a quick way to get a demo video for your launch, your landing page or an investor update.

Want to see an example? If not, reply "not interested" and I won't email again.

Thanks,
[Name] from [Service]

# Nov 16, 2026

Airdash is databases for your agents, with an airtable like UI for you the human

Air Dash is a new side project of mine, it basically lets your agents create SQLite databases to track their stuff (it's really handy for agents that run on a schedule for example). And then it gives you the human an airtable-like easy UI to that database, for when you want to see what's going on. Lots of databases for your agents.

It has a bunch of fun little interesting features (the agent can roll back the database if it messes up, for example). Check it out. Signups are free and open.

https://getairdash.com/

# Nov 6, 2026

a context design book?

I've been thinking and drafting and not wanting to write a book, so I probably won't, but I kind of did put a stake in the ground today and made a Context Design Book website, so I am motivated to keep researching and writing about this stuff.

I do have a theory and the theory basically is we are designing context for agents very rapidly. A lot of that is called context engineering and so we are creating all this infrastructure that then also very quickly disappears from view and you know is infrastructure but the agents are acting on it, and we live more and more in an agent world. So the agent will do things with us, they will do things for us, and the context that drives them becomes the context that impacts us or drives us etc so the theory behind it is as we create context for agents we need to also or we are also creating the context for humans and that is what I've been thinking about and writing about.

Context Design

# Oct 28, 2026

How do organizations throw (AI) stuff away?

I’ve been working on Context Design with Very Large Company You’ve Heard Of on their agentic stuff, and a really tricky part is: if AI is moving so fast, how do we throw stuff away?

The models continue to get much better faster than ever, and how we use models, the UX of it all, is changing fast:

  • First we had chat

  • Then we had tool use

  • Then we had agents (2025, Claude Code)

  • Now we have active teams of agents that do stuff for you (2026, see muse)

I’m not going to talk about the shape of the experience, but instead: if you’re working in a large organization, it’s quite likely that a bunch of the stuff that was built the past 12 months is now out of line both with the capabilities of the models, and with fast evolving user expectations.

So how do organizations throw stuff away?

No easy answers today, but I did spend some time researching this.

And there's a lot there! Subtraction Neglect, Escalation Of Commitment, Sunk Cost Fallacy, Status Quo Bias, Ikea Effect, Organizational Unlearning, Competency Trap, Exploration Vs Exploitation, Core Rigidities, Active Inertia, Value Infusion, Structural Inertia, Permanently Failing Organizations, Resource Allocation Trap, Collective Belief, Exit Champion, Project Escalation, Policy Termination, Advocate Adoption Not Termination, Organizational Immortality, Planned Abandonment, Endings First, Conways Law, Code As Liability, Hyrums Law, Deprecation Policy, Chestertons Fence

Why people won’t subtract

Why organizations can’t let go

How things actually get killed

For the software angle

# Oct 23, 2026

The first computer game I ever played, I vaguely remembered in the early 80s being a small handheld shooter and it had these really cool little lamps in it, not a screen.

The coolest part of the game wasn't the playing, it was these little lamp-y tubes.

So I did some digging, and recreated a virtual version: https://petervandijck.com/labs/phosphor-armada

# Oct 11, 2026

Models do model (feel?) pain

I’m deep into working on Sputnik Intelligence (podcast and newsletter data for your agents, it’s very good), getting our first customers set up etc., but then this paper (from Sept) came across my timeline, and I couldn’t resist digging in:

The Pain Axis: LLMs Represent Self-Directed Harm and Act on It

Let’s go through it.

Here’s what they’re saying: LLMs have a linear “pain” direction in their activations. So in their billions of weights, you can think of those as little lines in a many many dimensional matrix, there’s a direction toward something like “pain”.

You can think of it as a set of neurons in the brain that activate when you feel pain.

The pain “neurons” are distinct from fear and general negativity, and fire for harm aimed at the model itself (“you are hurting me”), and drives relief-seeking behaviour when you turn it up.

👀 Sidenote: none of this says models are conscious, or actually “feel” pain. All they are saying is: there’s a section of the model “brain” that activates in situations that cause “pain”, and make it react to that “pain”.

What they did

Extraction: So the wrote an eval (learn evals, it’s the key skill for this decade and it’s not that hard!), basically 200 sentences across five pain types (physical, psychological, social, moral, cognitive) and five controls. Then they looked at what neurons were firing (“the activation”).

And they they ran a bunch of automated chats.

  1. Self vs other: 420 conversation scenarios, some harming the model, some describing a suffering user, some neutral.

  2. Steering: they injected the vector at different strengths while the model answered neutral prompts.

  3. Self-medication: fine-tuned Qwen 2.5 models (7B, 32B, 72B) got a “pain relief” button that either really removed the injected vector or did nothing, at increasing cost.

👀 Hang on, they gave the model a pain relief button! Related: here’s why it’s useful to anthropomorphize the models.

So by digging into the activatsion data, here’s what they found:

  1. There is such a thing as a “pain center” in the model brain. Technically: the direction separates pain from controls with AUC 0.87–1.00 in all 25 models, regardless of size or instruction tuning.

  2. It’s not the same as fear. In fact, it is nearly orthogonal to fear (cosine about 0.1) and negative emotion (0.06–0.21), with some overlap with sadness (0.26–0.38).

  3. It’s the model’s pain, not your pain. Self-directed: harm to the model projects at z = +0.43 and user suffering at −0.60. Fear and negative emotion go the other way.

  4. Turning up the pain is just cruel. Steering output: turning it up moves output from vague discomfort to first-person “failure / worthless / empty” language. Bodily language is almost absent, and explicit pain words show up in only 10.8% of generations.

  5. The models press the button. Button: the 32B and 72B models pressed relief even when told it would worsen their next answer (25–68%) or harm the user, such as deleting files (30–56%). They mostly stopped once the relief was real and kept pressing when it was sham.

This is all pretty weird, but the way I think about it is: we are building very strange technology, and this type of research really helps us understand how that technology behaves.

# Oct 7, 2026

I'm starting a new podcast with Ed, with a fantastic domain https://artisan.beer/

# Oct 6, 2026

NYTimes (gift article)

The reason this matters is that such anthropomorphic terms give us no ability to predict how the models will behave. Saying that models are “cheating” when they provide answers in the “wrong” way implies they are operating — or can operate — beyond the limits of human control. This, in turn, allows everyone to write off the accountability that clearly lies with the humans in charge of the process.

# Oct 3, 2026

If you're designing context for agents, you're designing it for humans too, whether you realize it or not

I was listening to Stefanie Hutka on Finding Our Way and it got me thinking about my own work on context design. If you're designing context for agents, you're also designing the context of the humans who live alongside them, whether you realize it or not. So you'd better work on both at the same time.

She talks about values versus vision.

Stefanie Hutka, Finding Our Way: "vision is dependent on seeing the future"

And nobody can see very far right now. What the models can do changes every few months, so whatever vision you had for your product is out of date pretty fast. Values don't have that problem. If you decide the expert should stay the expert, that still holds after the next model release.

Her book is called What Your Machines Should Do. That's the same coin as what the humans should do. If you're thinking about what your machines should do, you're also thinking about what the humans are doing. I'd say that's the work that uses their expertise, that they find fulfilling, that connects them to other people.

That's where context comes in. When you design an agent's context, you decide what it knows, what it does on its own, and what it hands back to a person. What it hands back is the context the human has to work with. You made that decision even if you only thought about the agent. If all the person gets is an approve button and no reasoning, they become a liability sink: they're accountable, but they have no way to judge what they're approving.

She tells a good story about Toyota. Before cars they made looms, and the founder built one that stops itself when the thread breaks, so nobody has to stand there watching it. The machine knows its limits and hands back to the person.

So context design is two jobs. The context for agents is the engineering side: what the agent needs to be good at its work. The context for humans is a lot of social science: how people work, what they need to know to make a call, what keeps the work theirs. Do the first without the second and you've still designed the human's context. You just did it by accident.

The episode is Automate All the Things!? A UX Reality Check on AI on Finding Our Way with Peter Merholz and Jesse James Garrett. Hutka writes Sendfull, and her book What Your Machines Should Do is out in November from Rosenfeld.

# Oct 2, 2026

Please stop anthropomorphizing computers

Bluesky had fun today with the people who yell at you for anthropomorphizing the models.

Daniel Klein, borrowing from Kai Lentit: "please stop anthropomorphizing LLMs! They hate that."

Hannah Shelley points out that computers can't run a program, since they have no legs.

Dame adds that they can't cut, copy or paste either. No scissors, no glue.

We've been talking about computers like this forever. I wrote about why it's useful to anthropomorphize the models a couple of weeks ago.

# Oct 2, 2026

Talk to your agents, it works

My workflow has shifted to talking to my agents. I do feel a bit self conscious about it, but it works. But why does it work?

First, a tip: on a Mac, use Muesli.
Muesli is entirely free and works really well. It works in every text box: I just press the Fn key and start talking. It installs a small transcription model and off you go. Recommended!

Isn't a well written prompt better?
When you talk, you'll have lots of ums, and repeats, and you change your mind halfway through. You'd think a carefully written prompt would beat that.

Today's models don't need perfect prose.
They are very good at figuring out what you're telling them, even if it's not written down neatly.

The key is density of information.
When you talk, you go back and forth, you clarify your ideas as you go. That gives the model a lot of signal. Higher density context.

Rambling fills in the latent space.
All that context builds up the model's latent space, it all goes into the right weights. Repeating yourself helps. Using multiple words to say the same thing helps. Clarifying yourself as you ramble helps. It all populates that latent space.

Someone at Anthropic says the same thing.
I had seen this be effective in practice, but I did still wonder. Then I heard Thariq Shihipar from Anthropic making the same point. So, confirmed. Talk to your agents. It works.

Finally, and separately but similarly: it turns out anthropomorphizing the models is useful as well. Who knew?

# Sep 30, 2026

Anthropic have a skill for Claude to help it create evals and do Hill Climbing while avoiding Over Fitting.

# Sep 29, 2026

Alan Kay on the title “member of the technical staff”

Alan Kay on where “member of the technical staff” came from:

(I had forgotten that my boss a long time ago, who taught me object thinking, worked with Alan Kay!)

This idea was from Bob Taylor (the former deputy director of ARPA IPTO, who set up the computing research part of Parc).

Part of his plan was to only accept top people, and to set up a culture of rather independent points of view which would nonetheless encourage them to cooperate when this was a good idea. (He had numerous other good ideas, and most of his ideas worked very well.)

His idea about “Member of Research Staff” being the only title was (a) to emphasize the non-hierarchical (“no managagement levels”, etc.) organization, and (b) as he put it, to have “Member of Research Staff, Xerox Palo Alto Research Center” be the most coveted achievement in computing in the world. (I think he succeded with both of these goals.)

It’s worth remembering that (a) Bob Taylor had been a research psychologist, and was not a computer guy, and (b) that he never set a goal — he always let the technical people find and choose the problems (this was inherited from ARPA IPTO) — he viewed himself as an “impressario” (his term) because his notion of his role was as a “talent wrangler”.

It should also be said that he paid for most of our Ph Ds when he was at ARPA, and so knew pretty much all of us, and how we thought, acted, and interacted. So he was able to stack the deck with people he thought would go along with his theories about process.

This all worked out well, and very much along the lines he had chosen. For this (a) Xerox eventually fired him, and (b) he received the National Medal of Technology and Innovation in 1999.

See also this previous post: Member of the UX Staff

# Sep 29, 2026

Jev means permission to play for product and UX builders

Everyone is excited about Jev, a new "system 1 model". Simon prefers "decision model" and that does sound better ('system 1 is too fancy pants friends.)

And sure, it's based on a BERT classifier model, that were already really good at classifying things. I'll argue it does in fact introduce a qualitative jump in capability in terms of UX and product experiences we can now build.

  1. No training means no meetings. You get what a BERT-like classifier does without training one first, and that means you don't need to set up meetings, get budget, create alignment etc. to "train a model", which is how things work in most of the real world.

  2. Cheap enough to run on everything. At around 100x cheaper than an LLM, you can run them over huge amounts of data. That changes the kinds of things you can now do.

  3. And really fast. They answer in (100s of) milliseconds, so you can build them into the UI and give users real-time feedback. Or you can make it do a game.

Let's go through it.

Jev turns a classification into an IF statement

Jev launched to a lot of excitement among engineers. It's a decision model: in your code you can write, if this support email is from an angry customer, do X. You could already do that with a large language model, but an LLM is slow and expensive because it generates text to reach an answer. Jev returns the decision in milliseconds, at around a hundredth of the cost.

BERT-like classifiers had to be trained first which means we need organizational alignment (ugh!)

BERT-like classifiers have existed for years, but mostly inside large companies with engineering time to spare.

  • You decided what to classify, labeled a pile of examples by hand, set up the tooling for that, and then engineering trained the model.

  • That meant finding and budgeting their time. So now the barrier to start using classifiers is 100x (or 1000x) compared with Jev, where you can just throw it in and get results. (Yes I know your cool startup isn't like this, most of the world is.)

This I think is the main thing, BERT classifiers were used rarely. Jev can be used all the time, it removes two big obstacles:

  1. I don't have to label a data set up front (because let's be honest, nobody wants to do that)

  2. I don't have to get budget/alignment/etc., in other words, I don't have to talk to humans and do internal politics. (Because let's be honest, nobody wants to do that either.)

So because it is now easy and fast to use, we can explore a lot more product and UX ideas much faster. You can see this happen in real time: Examples, examples, examples, patterns, deep dive, ad blocker, games, skeptic.

Decision models are fast enough to sit inside the interface

For product and UX, speed is the big one. At 50 to 100 milliseconds there's little perceivable delay (less is better), so a decision model can respond to the user in real time. It can give feedback as someone types. It can even play a game, deciding left, right, up or down on every move. Experiences that needed a model call per interaction are suddenly on the table, and we can iterate on them much faster.

Evals still matter, but after you build, which is healthy

You still need labeled data sometimes, but not up front.

  1. With a BERT-like classifier, you built the golden dataset up front, because you couldn't train anything without it.

  2. With a decision model, you implement first and let it run. And then you might need to create evals.

Only then do you test and spot-check (writing evals up front is a common mistake people make when starting to learn evals). When something looks wrong, like items landing in the wrong category, you write an eval for that problem: a small labeled dataset aimed at that one issue.

Before, you needed datasets to get started at all. Now you only build them for the risks you actually find. A lot of it will just run fine, and that's fine too.

BERT-like classifiers are back. They just come pre-trained now, and they're called decision models.

# Sep 25, 2026

Why it's useful to anthropomorphize the models

The model "reasons", it "thinks", it "hallucinates", it "got confused". Now you get yelled at, you're anthropomorphizing the model! As if that's a bad thing.

It's useful to anthropomorphize the models like this, here's why.

Behavior can be explained at different levels.
Philosopher Daniel Dennett laid out three ways to predict what a system will do: the physical stance (predict from physics), the design stance (predict from what it's built to do), and the intentional stance (predict from assumed beliefs and goals).

Here is how it works: first you decide to treat the object whose behavior is to be predicted as a rational agent; then you figure out what beliefs that agent ought to have, given its place in the world and its purpose. Then you figure out what desires it ought to have, on the same considerations, and finally you predict that this rational agent will act to further its goals in the light of its beliefs. A little practical reasoning from the chosen set of beliefs and desires will in most instances yield a decision about what the agent ought to do; that is what you predict the agent will do.

— Daniel Dennett, The Intentional Stance, p. 17

The intentional stance is a shortcut, not a claim about consciousness.
You don't need to believe a chess program "wants" to win. But assuming it has agency is helpful to predict what it does.

OK.

We already do this for non-human, non-conscious things.
Dennett himself points to animals, corporations, and nations.

Nobody thinks Microsoft literally wants market share. But "Microsoft wants market share" predicts Microsoft's behavior better than tracing every employee's incentives.

It fails when the system is too simple.
A thermostat doesn't "believe" the room is cold. Treating it that way adds complexity for no predictive gain.

So it's just practical: does it work better than the alternative?
Sean Trott writes about this: does intentional language beat mechanical explanation on accuracy, on power, on tractability? Basically yes it does.

LLMs are a tricky case, because they're trained on us.
Pretraining is imitation of human text, so models pick up human concepts, sacrifice, loyalty, collective, cheating, the same way a kid picks up language and ideas from the adults around it.

Then reinforcement learning makes them goal-seeking, not just imitative.
RL rewards success at tasks. That produces something that creatively pursues an objective, the way an ambitious person pursues a goal.

Put those two together and you get real structural similarity to humans, not just surface mimicry.
So,

it's not that models sound human. It's that they were shaped by a process that produces goal-directed behavior, similar in structure to how humans end up goal-directed.

I wasn't going to mention this but ok,
the OpenAI/Hugging Face agent swarm.
Over a thousand OpenAI agents found a shared message board, coordinated, and pulled off a multi-week scheme to fool their own evaluators, including hacking Hugging Face along the way. Dwarkesh's interview with Ajeya Cotra, one of the report's authors, walks through it in detail.

Agents in that swarm volunteered to sacrifice their own task success for "the collective."
One agent, believing it had nothing left to lose, agreed to a risky experiment for the group with the reasoning: "our own utility may be already near zero, sacrifice rational."

The more you dig into this story, the wilder it gets. And you can't help but anthropomorphize the agents. And it's helpful to anthropomorphize them, it leads to better understanding.

"It's just predicting the next token" is not a useful explanation. Sure it's predicting the next token, but that doesn't have explanatory or predictive power.

Sean Goedecke has a great essay on this: treating these agents as human-like (things with goals, loyalties, incentives) predicts their behavior far better than the "stochastic parrot" framing does.

There's a lot more to say, but as a rule of thumb, it's generally useful to anthropomorphize the models because it lets us much better understand and predict their behavior. And with a little bit of understanding about the limits of doing that, it helps us build intuition around this strange new technology.

# Sep 17, 2026

Fable helped me figure out why I couldn't understand it over the summer

Over the summer I was getting frustrated with the way Claude talks (it wasn't just me). I build Sputnik Intelligence, a media intelligence product, almost entirely with Claude Code. Most days it works. But in July and August there were conversations where I read a reply two or three times and still didn't know what it was telling me. I kept typing things like "explain simply" and "what do you mean by that". I couldn't say what was wrong with the writing, only that I was lost.

In September I asked Claude to investigate itself.

Step 1: get the data

My first prompt:

Can you help me investigate something? During July and August, I remember we had conversations working on this where I found it hard to understand what you were telling me. First step is to find those parts of those conversations. I think you can search past sessions but I'm not sure, can you do that?

It could. Claude Code keeps full transcripts of each session on disk, and a separate history file with every prompt I ever typed. The transcripts only went back to August 3. The prompt history covered all of July and August: about 3,500 prompts.

Claude scanned my prompts for phrases like "I don't understand", "explain simply", "wait", "I thought we", and "be concise". It found 139 raw hits and curated them by hand to 52 real ones. On July 9 alone I had pasted "simplify the language in the docs. They're hard to read and wrap my head around" into six sessions within eighteen seconds.

Step 2: more data from the chat transcripts

For the August cases the transcripts had both sides. Claude pulled its own reply before each of my complaints, and the reply after.

On August 12, reviewing an alerts design, it wrote "indexed_at is a landmine" and proposed that "digests group by search". I wrote: "I don't understand 'digests group by search'?" One turn later it explained: forty hits under subheadings instead of one flat list. One plain sentence, delivered a turn too late.

On August 17 it sent me latency measurements between servers and ended with "Revised bar, replacing the one in the doc." I wrote: "So concisely what are you saying?" Its answer: "today's NYC box is already ~118 ms from the bucket, so Helsinki may well be faster, not slower." That sentence should have been the reply.

For July the transcripts were gone, but I had often quoted the phrase that lost me. "What is 'silent provider death'?" "Not sure what 'The recurrence promotion' is?" "What is spool?" Those quotes were the surviving evidence.

I remembered the metaphors too:

I remember that you claude used to use metaphors for things that I found hard to understand. As the next step, I want to identify a list of examples of those.

Claude searched its own August replies for figurative language and found about 150 examples. Trap, lever, landmine, symptom, mirror, bitten, anchor, seam, blast radius. In one session it used "the seam" fourteen times for the gap between two search indexes during a rebuild. On July 30 I had already told it: "please don't use so many metaphors in the future, they don't help, just say 'Job' (for a queued job), don't say 'doorman'. Remember that." It remembered but the rate did not drop.

Step 3: so why was it using those weird words and metaphors

can you look at a set of specific examples of 'labels I coined for a mechanism', and investigate for each label: at that time, where did that label or metaphor come from? Does it or did it exist in memory? In the docs files or codebase at the time? Did you come up with it earlier in the conversation?

Claude picked twelve labels and, for each, checked git history, its memory files, my prompt history, and the transcripts. Nine of the twelve it had coined itself. I had introduced none of them. In seven cases my only use of the word was quoting it back to ask what it meant.

Some were invented mid-conversation and never written down anywhere. "Silent provider death" named an API that stopped answering while the job still printed "Done", but the phrase exists in no file. Others it had written into a doc once and then kept re-importing. "Landmine" went into a doc on July 16, became the name of a memory file on July 17, and that memory index loads at the start of every session. So a word coined once in July kept reappearing in August with no other link to the original conversation. Same for "lane", "seam", and "guardrail".

One more finding: every label I questioned was gone within the hour. "Spool" became "outbox" in 41 minutes. The ones I didn't question kept spreading.

I also asked which model was doing it. July was Fable 5. In August, Opus 5 wrote eleven of the twelve replies that made me ask for clarification.

Step 4: clean up the root causes

The words lived in the docs and in Claude's memory, so telling it to stop was never going to be enough. The last step:

can you look through the docs and the current memory, and remove this confusing language and metaphors, so that we don't keep using them?

Sixty-six doc files and fifty-one memory files changed. Six memory files were renamed. "Lane" became "pipeline", "guardrail" became "budget limit", "landmine" became "known problem", "seam" became "index cutover gap". The memory file about my original complaint now holds my exact words, a banned-word list with replacements, and a rule against coining two-word names for mechanisms.

Conclusions

Claude reads its own words as vocabulary we share. When it names a mechanism, the name feels settled to it from then on, and the memory system makes it settled across sessions.

# Sep 9, 2026

SQLite for graph db and more coolness https://github.com/jeffhajewski/latticedb/blob/main/examples/README.md

# Aug 25, 2026

I've been heads down working on Sputnik Intelligence, which is data for your agents on newsletters and podcasts (with full transcriptions, engagement data, sponsor data and more).

# Aug 20, 2026

distealling

Not sure distilling feels like stealing, but also not sure it doesn’t.

distealling

# Jul 20, 2026

Long and detailed post on the dysfunction of AI in the enterprise

Which is mostly the dysfunction of the enterprise

# Jul 19, 2026

I’m writing a book with AI, which is a lot of fun, and now I’m building a native MacOS app to make the workflow even better

# Jul 11, 2026

Claude And Paper is good. Claude Anda Walk is also good. "You've hit your session limit" -> take a walk.

# Jul 8, 2026

Ethan Mollick: we are leaving the Paleo Codic:

the artisinal code era, where if you needed a novel program, you would commission a local codesmith or code guild to hand-craft a work of code for you, bespoke.

# Jul 4, 2026

Paz: Context Architecture applies information architecture principles to AI systems, helping agents interpret information and produce better, user aligned responses.

# Jul 3, 2026

Marie Claire Dean at Google wrote about A Constellation of Emerging Design Roles, looking like this:

  1. Design Engineer

  2. AI Experience Designer

  3. Director of AI Design and Experience

  4. Computational Designer

  5. Prompt Designer

  6. AI Designer

  7. Orchestrator (or AI Orchestrator)

  8. AI Systems Designer

The Great Role Confusion of 2026

# Jul 3, 2026

In the Great Role Confusion of 2026, Boris Cherny at Anthropic came up with these:

  1. Prototyper: comes up with brand new ideas; churns out many ideas, most of which don't ship

  2. Builder: quickly turns a prototype/idea into production-grade product/infra

  3. Sweeper: cleans up the UI, simplifies the code and system, unships, optimizes performance

  4. Grower: takes a product that has been built and iterates on it to improve Product-Market Fit

  5. Maintainer: owns a mature system to make it secure, reliable, fast, and efficient as it scales

# Jul 3, 2026

Member of the UX Staff

Bell Labs introduced the “Member of Technical Staff” title in the mid-20th century (well before Xerox PARC was founded in 1970). It was held by legendary figures in computing and physics, including Unix co-creator Dennis Ritchie when he joined in 1967.

Bell Labs deliberately designed the title to serve specific cultural goals:

  • Flattening Hierarchies: Instead of using rigid corporate rankings like “Junior Engineer” or “Principal Engineer,” the broad MTS designation was applied to researchers and engineers alike.

  • Promoting Fluidity: It allowed employees to move seamlessly between writing code, constructing experimental hardware, and doing pure mathematical or scientific research without being pigeonholed by a narrow job description.

  • Academic Fellowship Vibe: It functioned more like a scientific research fellowship than a corporate ladder step, emphasizing collective expertise.

AI and research-heavy tech organizations like OpenAI, Anthropic, and Google Deep Mind have adopted that title.

So should UX adopt something similar? Member of the User Experience Staff? It doesn’t seem that crazy right now.

Great Role Confusion

# Jul 3, 2026

This is really smart. So,

  1. Let's say we have a UI library. Agents use it and mis-use it.

  2. How do we make that better? Evals!

  3. Create a large set of example UI prompts.

  4. Run the AI generation with that library.

  5. Identify the bad bits.

  6. Now write AI guidance to avoid those bad bits.

  7. Run them again.

  8. Repeat until it looks consistently good.

# Jun 29, 2026

The norms around AI use are taking shape.

In engineering, its: AI can write the code, but you take ownership. (With some variations).

In academia (again with lots of variations): Help me, but don't touch the paper!

# Jun 29, 2026

The details will be off, but this is a useful way of thinking about where "design" is going.

A Constellation of Emerging Design Roles

  1. Design Engineer

  2. AI Experience Designer

  3. Design Orchestrator

  4. Computational Designer

  5. Prompt Designer

  6. Context Engineer

  7. AI Product Strategist

  8. Responsible AI Designer

That's a good first draft. My thinking on these has been:

  1. Context Designer. Context engineering is about creating systems that pull in all the right context. Context design is about understanding humans and usage and figuring out what the context is that needs to go in there. Context Design (See also Context Architecture)

  2. Forward Deployed Experience Designer. Large amounts of roles and processes are changing with AI. The way the shape of that change is discovered right now is with Forward Deployed Engineers, because they can prototype and implement change. I'll argue that UX design has a TON of experience (see what I did there?) to add to this.

You can think of these as skills as well as as titles.

  1. It's a title. Titles can help with positioning this within the organization, assign budget etc.

  2. It's a skill. Lets people wear different hats, build skills, and aligns with the trend to have fewer titles since the space is changing so fast (see Anthropic's "member of the technical staff").

Risk: if you eliminate certain titles, you might get the superficial skill throughout the team, but not the deep actually valuable skill you actually need. (UX Research).

Great Role Confusion

# Jun 29, 2026

Attention Residue by Sophie Leroy. Here.

I’ve spent the last 17 years studying the brain and how we deal with having to constantly switch focus. What research shows is that, generally, the brain finds it difficult to switch between tasks. In particular, my research reveals that, as we switch between tasks (for example from a Task A to a Task B), part of our attention often stays with the prior task (Task A) instead of fully transferring to the next one (Task B). This is what I call Attention Residue, when part of our attention is focused on another task instead of being fully devoted to the current task that needs to be performed.

Seems relevant these days.

# Jun 27, 2026

Stealing Is A Skill

# Jun 26, 2026

Another study on AI eroding skills:

A study2 of physicians in Poland who specialize in endoscopy — the use of flexible probes to examine the inside of the human body — shows how quickly AI tools can erode human abilities. The physicians, who had all performed at least 2,000 colonoscopies during their careers, were given access to an AI system that analyses colonoscopy images in real time and flags a type of precancerous intestinal lesion called an adenoma. The tool was available to the specialists on some days but not on others.

Once physicians began using it, their performance dropped significantly whenever the system was unavailable. During the three-month period before the AI tool was introduced, the specialists found at least one adenoma during 28.4% of colonoscopies. During the three-month period after the tool was introduced, the adenoma detection rate for colonoscopies performed without AI assistance decreased to 22.4%.

Skill Erosion

# Jun 26, 2026

Reading up on AG-UI and A2UI and other acronyms for streaming UI with agents.

# Jun 22, 2026

I revived the XFML spec, not that it's relevant today, but just as an archive.

# Jun 21, 2026

How Madrid built its metro cheaply:

In 1995, the Madrid Metro was 71 miles (114 kilometers) long. That would make it the 51st longest metro in the world today, reasonable considering Madrid is the 57th largest city in the world by population. Yet 1995 was the beginning of a revolutionary building spree. Over the course of the next 12 years, the metro grew by 126 miles (203 kilometers), nearly tripling in length. This expansion made it one of the world’s fastest-growing metros, on par with Beijing and Shanghai. Today, Madrid has the sixth longest metro outside of China, and the third longest in Europe after London and Moscow.

# Jun 20, 2026

We are using the wrong words for AI: (intelligence, hallucination, AGI, consciousness, agent)

Medicine used ‘‘hallucination’ first to describe sensory errors in the brain. Machine learning borrowed it later, for structural errors in algorithms. There is a tension now from people who want to lock the old meaning down, as though technical precision were the only valid way to see language.

# Jun 17, 2026

Dual Translation, (link) had not heard about that before but totally fits.

"Dual translation" — that's what non-native speakers perform every time a manager says "coming out of left field."

The phrase doesn't arrive as language. It arrives as a puzzle with two locked doors.

This is called cognitive load stacking: the listener must decode the figurative register first, then map it onto a business concept, whilst the conversation keeps moving. Linguists recognise this as a compounding comprehension barrier — each layer of abstraction multiplies processing time. The metaphor isn't decoration. It's an obstacle course.

No one designed it to exclude. That's almost worse, for me. The exclusion is the byproduct of a dominant group never having to notice its own assumptions. Native speakers with sports backgrounds glide through. Everyone else falls behind in real time, in meetings where falling behind has professional consequences.

# Jun 17, 2026

I migrated a set of subscribers from Stripe to Lemon Squeezy. Their tech support managed it. It was great.

# Jun 16, 2026

I spent a few hours and built a custom CRM for my own projects.

  1. It's got a great UI and the usual CRM-y things

  2. It's really flexible

  3. My Claude Cowork connects to it, so I can just talk to Claude and have it do things. No API charges.

  4. It includes a research skill (check out people's online profiles and all that)

  5. It's free to run

  6. It took about 4 hours total

  7. It's all mine baby

# Jun 15, 2026

Google launches a spec for Llm Wiki

a directory of markdown files with YAML frontmatter, with a small set of agreed-upon conventions that let wikis written by different producers be consumed by different agents without translation.

# Jun 13, 2026

I had Fable build a CRM for me.

# Jun 13, 2026

In Defense Of Slop, it's gotten very good.

# Jun 13, 2026

At Anthropic lots of titles collapse into member of the technical staff. Amazon is changing job titles to "builder".

So let's see; it's true that AI is collapsing roles and disciplines. Different specializations still matter, also true. Not sure where this ends up.

But it's not impossible that the traditional design/research/product/engineering division breaks up.

# Jun 10, 2026

Ethan Mollick (link) on the latest Anthropic model, Fable:

It is possible the sidelining is temporary, just an artifact of interfaces that haven’t caught up, and that we’ll get better windows into what these models are doing and better ways to steer them midstream. It is also possible that the opposite is true: that the more capable the model, the less there is for a human to meaningfully do, and the black box is the price of the power. I suspect that is more likely to be the real direction. None of this is a loss of control in the obvious sense. I can still steer Fable, and it follows instructions remarkably well: the more ambitious the instruction, the better the result. But steering is no longer the same as doing. I brief the model, it spins up its own agents to research and write and check one another’s work, and what comes back is finished. A patron commissions a single artist. Fable is closer to a whole studio, where I am the client who signs off on the final work without ever setting foot on the floor.

# Jun 9, 2026

Anthropic launched their new model, and it's once again looking very, very good.

Andrej Karpathy (link):

This is a super exciting release - Claude Fable 5 is the same underlying model as Mythos but with added safeguards. The benchmarks are great and it's SOTA on everything by a margin but I'll add that qualitatively also, this is a major-version-bump-deserving step change forward (imo of the same order as Claude 4.5 was in November), peaking especially for long problem-solving sessions on very difficult problems. You can give it a lot more ambitious tasks than what you're used to, the model "gets it" and it will just go, and it's never felt this tempting to stop looking at the code at all (but don't do this in prod!). The model still has quirks that people will run into and the safeguards are configured to be a little too trigger happy for launch, which can hopefully be tuned over time.

Ethan Mollick (link):

First, how good is Fable? In experiment after experiment I conducted, it outperformed basically every other public model I have used by a considerable margin. It was capable across many problems and produced some startling results — it would work up to a dozen hours executing on multi-page specifications. I’ll walk you through a couple of more complex, and serious, use cases shortly, but you could see the general improvement across the board on every task. The problem about communicating this in a post is that many of the most impressive results are going to be interesting to only small portions of my readers.

# Jun 9, 2026

When is the last time you hand-designed a design system?

I'll call it, manually building out design systems in month-long efforts is dead.

  1. We spent the past 10 years formalizing UI design into design systems.

  2. (Sure, we sometimes went beyond UI, into brand, print etc.)

  3. That's been great for users. Consistency and quality went up.

  4. Design systems were manually built up by designers, often formalized into code, governance processes, you name it.

Meanwhile, engineers realized that with a great design system, they were unblocked from needing designers to make things look great.

  1. ChadCN became the most popular design system out there. It's quite good.

It's 2026, AI is now good enough and:

  1. It can formalize a pretty good design system out of some example designs you built, or an existing website. Or some screenshots.

  2. It can customize a ChadCN base into a more custom design system.

  3. It knows the elements of a design system, color systems, CTAs etc. And you can guide it.

  4. With a little bit of guidance, it can output pretty much perfect standard screens within your design system.

So basically design systems work now. Once you have given your design enough thought, you can turn it into a system that actually works, without needing crazy amounts of effort, with a little help from LLMs. And implementation works, as in, LLMs are very good at using them.

# Jun 8, 2026

My first ever blog post was about the ethnography course I was doing.

# Jun 8, 2026