adriand

None of these mitigations (although they’re good ideas) address what I think is the real issue here, which is that people don’t like AI. And why would they, when the dish being served by AI CEOs is job loss with a side of extinction risk.

I think many data centre opponents understand that being opposed to AI is not sufficient to halt a local project, so it is noise, water use and power consumption that they latch onto - although, in jurisdictions without plentiful water and clean power, those two reasons are quite valid.

In my own local community, there’s vocal opposition to a proposed data centre that is next to Lake Ontario, 500 metres from the nearest residential neighbourhood, on a brownfield site with 500 megawatts of clean power capacity. None of that makes a difference to those opposed.

What would you like to have instead?

History! That's what intrigues me the most: texts with accounts of events that have otherwise vanished from the historical record.

What are the wildest, most exciting but plausible things that might be discovered in these documents?

Go look at Messenger on the App Store and look at the “data linked to you” section. An app with that much surveillance baked in is surely not improving the world.

I think you are correct about the legislative push, because it’s clear that US AI companies can no longer live without it. However, it is not the case that this is Amodei’s fault. His push for regulation was clearly at least partly mitigated by a desire for this precise situation to be avoided! With an appropriate regulatory framework and a transparent, apolitical certification or review process, this kind of situation would not happen. Banning Fable is not “regulation”, it’s capricious retribution, and I believe it is the single most damaging thing that has happened to the US AI industry to date.

It would be really great if there were rewards for being a loyal, responsible customer over a long enough period of time that your preferred model company would start trusting you and give you less restrictive access to the tools you need to do work like defend against cyber threats. I noticed recently that after a year or so, Stripe now lets me do “instant payouts”, presumably because I now have a track record of responsible behaviour. AWS also does similar things, especially for things with abuse potential like SES.

I would really like to live in a world where the “good guys” have terrific tools and defenses at their disposal. Instead it seems like we are heading for a world of empowered bad actors and hobbled ordinary citizens.

On the plus side, it’s Friday night. Hopefully this is sorted out by Monday morning.

I’m still on Sequoia because of all the Tahoe complaints. Not so much the rounded corners, but I heard performance was poor. Have updates to Tahoe improved things?

Wow - the writing in this is fantastic. This is genuinely hilarious!

What kind of reputation does ca-central-1 have? I’ve been using it and it seems quietly excellent. Knock on wood.

I think so, but that's also really great because I frequently run into the five hour caps, but very rarely use my entire weekly allotment. There are lots of situations where I do things like write the plan for all the work that has to get done, and then set a reminder to execute the plan after I get home, when I'm done making dinner (because e.g. my five hour cap ends at 6pm). Higher caps for the five hour period is a lot more convenient.

I went to Newfoundland and I went to a bar one night and met a guy from a small town along the coast and I literally couldn’t understand a single thing he said. He was apparently speaking English but it may have been Ancient Greek for all I was able to make out. The only way we could communicate was via the bartender, who would interpret what he said and tell me. He had no trouble understanding me. It kinda blew my mind.

Kind of a great video! I enjoyed it. His point about testing coverage and generating mutations to ensure the tests fail resonated. I get concerned sometimes that the AI is writing tests not to ensure the logic is correct, but to ensure the tests pass against the code it already wrote. Any other ideas on this? Is there a code review step or CI checkpoint that would decrease the likelihood of that?

The narrative that they have guards against mentioning openclaw doesn’t make sense to me - I’ve been using Claude code to manage an openclaw instance for a few weeks now, with zero issues.

But wouldn’t that be true for all periods in the dataset? You see ups and downs over the centuries, and in each of those centuries I’m sure humans heavily influenced their evolution. Then you see a pronounced upward trend that just happens to also coincide with what we know to be serious, sustained and highly unusual planetary-wide warming.

But would you want the AfD to come to power and wield those ramped up, potentially now nuclear, forces?

Totally! That’s what makes the situation doubly maddening. It would be one thing if these actions were bad for the world and good for the US. But they’re bad for the US too!

I forget who it was that said this, and I’m sure my paraphrasing is bad, but I listened or read something I found chilling. It was something like, ordinary Americans are totally unprepared for the level of danger they will experience over the coming decades.

The only reason Trump is able to destroy global institutions so easily is because Americans take their security for granted. But that security is the result of institutions developed in the aftermath of an utterly devastating war. Now those institutions are damaged and America’s friends are alienated, right when they are most needed to deal with China, Russia, AI, drones, cyber, nuclear, climate…talk about bad timing.

The US on the other hand is still there with much larger force.

To provide for European security! That’s the deal in terms of Europe and NATO and also specifically to handle Germany. The idea was that America would provide security to Europe including the nuclear umbrella, and one benefit - among many others - was that Germany would not need to have a powerful military.

Can you perhaps guess why people might be concerned about a heavily armed Germany in the postwar period? Those same concerns are bubbling up in European capitals right now, as Germany rearms due to the loss of the US as a reliable partner.

The US did not retreat. We fought multiple wars to maintain our power and influence. We toppled multiple regimes to maintain puppet governments. Very much the same as the USSR and China have done.

As much as I am critical of the US, until now the US did behave very differently from other superpowers. Consider the end of WWII. The US did not inflict reparations on the vanquished nations but rather, invested huge sums in their rebuilding, in the process making stalwart allies of them. These were not puppet governments, they became thriving democracies.

This is not to excuse the many bad things the US has done in Latin America, Vietnam, etc. But there is really no comparison between US behaviour and that of the USSR (or of colonial European countries, for that matter). People in Soviet-controlled East Germany were quite keen to go to the west and did not perceive the presence of US military bases there as evidence of American totalitarianism.

That, of course, has changed and now America is seen as a predatory hegemon. But that has not always been true.

Who do you think suffers when elites attack public education? It's always the children.

Exactly. And who benefits from a less educated, less aware populace? The answer is pretty clear: look at who is benefiting right now!

In other words "ohshittification."

Brilliant coinage, if it’s yours, congrats!

My take: it is not enshittification to raise the price for a product whose demand outstrips its supply. That is basic economics. There are alternatives, it’s not a monopoly. If you think it’s the best product, then pay more for it.

Personally I would be perfectly content if the price of Max went up a bit and Pro no longer worked for CC if it meant that Max was faster and more stable.

Iterative experience (experimenting with different ideas, deciding what works best) and speed of execution (once I was happy with it, making it happen required almost no work).

I had a similar experience with running out of usage quite quickly, after setting up one design system properly, and then getting pretty close with a second one. But it's a research preview - I'm sure it will change.

I was quite happy with what I pulled off using the first design system: I wanted a new footer section for my IPAAS startup, it generated four options, the fourth of which was quite good. We iterated on it for a bit, then I pulled it into Claude Code (that integrated feature is very cool), CC built it, I deployed it, done. (Bottom section of https://tediware.com/ if you're interested, the bit with "Origin story" on the left and the signup panel on the right).

It was not a complicated build by any means but I liked the concept it developed and it was dead-easy to make it all happen. I think the ideas in the UI are very good. Still rough, but you can see where this could go, and it's got a ton of potential.

Yeah I really want more information than "on a walk". Really? No digging whatsoever involved? Did they walk past an eroding riverbank or something? I'm so curious.

There are still SO MANY insanely ugly, hard-to-use, absolutely horrible apps out there though. Sure, in consumer-focused apps, there's a lot of competition and pretty much everything popular is well-designed. But in enterprise? My god, it's still a massive shitshow.

The hilarious thing is that I would be willing to bet than in a decade, it's STILL a massive shitshow in enterprise. That's because the problem with enterprise software is not that good design is all that difficult to pull off (it just requires caring!) It's that the people making enterprise software have terrible taste and can't even see (I am convinced) that the thing they built is ugly and hard-to-use.

You can set it to auto top up if it drops below a certain amount. If you do that, then it would definitely be wise to add a cap. They let you add daily/weekly caps, which is convenient.

Anthropic doesn't have the compute so everyone can switch to Claude for a couple months, get nerfed, switch back.

This seems to be the new narrative around here but it's not jiving with what I'm experiencing. Obviously Anthropic's uptime stats are terrible but when it's up, it's excellent (and I personally haven't had any issues with uptime this week, although my earlier-in-the-week usage was lighter than usual).

I'm loving 4.7. I was loving 4.6 too. I use Codex to get code reviews done on Claude-generated code but have no interest in using it as my daily driver.

You mean openrouter.ai. And yes, on reading this blog post, I immediately reviewed my API keys in OpenRouter to make sure that they were capped. My prod key was capped at $20/day (phew!) but my dev key had no cap, which I just updated. What a horrible story.

How does that actually show up in practice for you?

It shows up as inconsistency. One of the key things I built in this architecture is the ability of users to define standard operating procedures (SOPs). These are the instructions (i.e. prompt) for the agent to do tasks (I've integrated Sonnet via OpenRouter into the SOP drafting UI, so people have help creating these - and the system prompt for this knows about the API endpoints that the agents have access to, so people get good advice as they write them).

Anyway, it's not uncommon for someone to write an SOP, test it a couple of times, decide that it works, and then tell people it's good to go. There's probably a 1 in 3 possibility that it doesn't actually work when someone else tries it. The reasons for that are almost endless it seems.

This is just one aspect. It seems like something new fails every day. Today:

- the agent stopped responding to incoming email. I dug into it. Somehow the tailscale hostname had changed. I had not changed it. I have no idea why it changed. This is not OpenClaw's fault, but it speaks to my point that there are too many moving parts with these things.

- the agent stopped sending emails when tasks were completed. This runs on a "cron" job. I went through the list of the cron jobs. The "task reporter" cron job was disabled. Why? No idea. I didn't disable it. I'm the sole "operator" of the OpenClaw instance, so if I didn't disable it, then something inside OpenClaw did. Why? I don't know.

What I do know is that someone pings me every day with a complaint that something is not working, which is a new experience for me, and it's embarrassing.

Is anyone enjoying the process of deploying agentic AI for clients/employers? I have been working on stuff in the space ever since a senior exec at a client became enamoured with OpenClaw and I had the choice of either letting her go full foot-gun or enabling her to do what she wanted with agentic AI. I built a fair bit of stuff that sounds like this project to try and put some guardrails, observability, security, etc., around OpenClaw, but I have to say that the flakiness of the overall system is a huge turnoff for me.

The lack of predictable output/outcomes, the number of things that can go wrong (rate limits, this or that service stopping, a "cron" job seemingly disabling itself, permissions that don't stick, on and on it goes), does not make for an enjoyable development experience. People are getting value from it and on some level it's quite remarkable what can be done, but never in my life have users of my software had such little faith that what worked properly yesterday will work properly today. I have had far better results using LLM APIs.

I always wonder if competitive market dynamics will solve problems like these, at least to some extent and for some people, because the people who retain the ability to communicate in a distinctive, persuasive and original style will be rewarded. Corporate dronespeak is no less homogeneous than AI writing, and companies with this communication style are regularly disrupted by nimbler, more authentic-sounding competitors.