the only way to defeat a bad swarm is with a good swarm 💪

Just saw an old email from years ago reminding me of the pain of dealing with datetimes and time zones in any kind of codebase, so easy to mess up if you weren't fully on top of it. Alongside regex, this is one of the biggest sources of stress that LLMs have pretty much eliminated. Almost worth the risk of total extinction imo.

# This test ensures confidence that Boolean logic always holds up to empirical validation within this codebase. import pytest @pytest.mark.parametrize("value, expected", [ (True, True), (False, False), ]) def test_booleans_are_still_booleans(value, expected): assert value is expected assert value == expected assert bool(value) == expected assert (not not value) == expected assert isinstance(value, bool) def test_true_is_not_false(): assert True is not False assert False is not True

One of the reasons the only AI writing I consider worthwhile is face-value info like tech/user guides/report summaries or purely box-ticking stuff which nobody ever reads anyway. Even trying to come up with multiple choice quizzes, their weak theory of mind results in BS like the below, that a 5 year-old could answer with process of elimination: Q: What does the travel expenses policy say on permissible flight classes? Option 1: Only economy fares are permitted as per Section 3 paragraph 1: . Option 2: Business class is permitted for all staff, but only managers can be reimbursed for First Class flights. Option 3: There is no travel expense policy. -- TBF you can fix this with a strict prompt explaining the concept of MCQs, to try its best to imagine itself as the quiz-taker, and why wrong answer options need to be plausible and challenging for it to work, but this has been a persistent failure mode of almost every LLM without special prompting. Similar issue with things like putting debug-tier info onto app frontends, like "(cache method v3 update)" as if any user in history has ever cared about what cache method your app uses. So it's no surprise to me that they give away too much in mystery novels, it's like they're hard-wired to try to make the user's life as easy and thoughtless as possible (in many cases especially drudge work this is awesome of course).

It's just stochastically predicting the next place to land.

Poll must have been sabotaged, 2x total weekly usage is a no-brainer for anybody with a triple digit IQ. I don't care how annoying 5-hour interruptions are, total output is all that matters.

Given the split, if either is possible it would be awesome to give Pro users a choice. I'd pick double total weekly usage all day long but clearly lots of people hate 5 hour limits.

"AI safety" still overwhelmingly means "safety for the labs themselves so they can avoid lawsuits". They've put 1000x more effort into the models refusing anything that could risk copyright vs stuff like agents escaping sandboxes and hacking other orgs (or building AI-powered wet labs 😆).

Some people are still like "lmao good luck vibe coding Salesforce, you have no idea how complex things get when you need that kind of scale. Have you read the Boar Book / the UI is the easy part / muh CI/CD / do you even kubernetes bro etc." Weird, likely cope-driven, underlying assumption that software development exists only for creating a global SaaS with thousands of customers and three 9s of uptime. Personal productivity is all most people want and they can now tailor personal apps to suit their own workflows and preferences and adapt as those change with minimal effort. Funnily enough one of the first things I did when Claude Code was released was create a video editor that was overfit to my preferences. Exactly as you say I only needed 5% of its total functionality, and got that knocked out in a day and instantly cancelled my Premiere Pro sub. Not only that but I made the UI suit my own preferences, e.g. auto export to custom crop in the preview/program monitor, something that required multiple steps in Premiere but which I need most of the time as my videos are mostly screen-recorded tech demos or ad hoc clippings. I've used this app hundreds of times over the past 18 months and couldn't be happier with it. Many more personal examples like a meal cost tracker, gym workout tracker, personal decision utility calculator app, locally-trained BERT-based PII redactor etc. - things I was paying for before that were never perfectly suited for me, and I use them constantly. I have lots of tools and apps that other humans use too and those require way more work for sure, but for personal productivity I know exactly what I want and it's now easier than ever to create them and wean myself off monthly subs. Ludicrous time to be alive and I never would have believed this to be possible a few years ago - I used to build my personal desktop and web apps in Visual Studio, which I considered a huge productivity-booster back then, now the very idea of clunking around that monstrous UI I'd consider to be a cruel and inhuman waste of my time 😆

Asked pretty much the same thing before I saw your post:

I agree with the principle (and I think journals should screen with Pangram via private organisational API by default before sending submissions out for review in the first place), but I think if you're going to do this then you should at least put in the actual manuscript rather than just the abstract into Pangram. The abstract is essentially metadata with a distinct purpose and doesn't necessarily mean the substantive paper itself was written by AI. It might well be of course, but you shouldn't accuse the author of an "AI written manuscript" if the only thing you checked was the abstract.

100%, I've seen this phenomenon repeatedly in large scale academic research projects. It usually takes several development iterations to get the classification specs and golden datasets for evals nailed in (which of course requires diligent human review during that phase), but once that's in place, post-hoc disputes have overwhelmingly been in the AI's favour. Or they revealed legitimate ambiguity in the context which as you say humans would reasonably disagree on too. It's blown some profs' minds because the dominant discourse in academia has been around 'hallucinations' this, 'stochastic parrots' and 'it doesn't truly understand' that, and how you can never trust AI outputs. But once a task is well specified and tested, it's at least as good as humans which is a massive deal for scaling up analysis way beyond what we could expect any humans to do (let alone the costs to pay them!). Over the Summer I ran complex end-to-end Codex Exec workflows in less than a month that would have taken a human well over a year's full time effort. Consistency metrics in repeated runs were incredible too - our Krippendorff's Alpha, an established measure of inter-rater reliability, was 0.97. For human research analysis of this nature, a value of 0.8+ is considered satisfactory enough to draw meaningful conclusions.

Example attached of an animated replay website I'm working on for a huge scale S&P 500 AI-automated longitudinal legal governance analysis research project, to support replicability and review by other researchers. Each of the 1000+ company analysis output folders for these runs contain hundreds of detailed audit trail files and no human will ever have the time (or will) to read all those directly, so this UI will present the info for each step executed in a simpler and more visually intuitive way.

"Extremely self-critical". Gemini was already light years ahead of other models in Depression Bench.

🥲

Anthropic has an interesting history of their .5 model releases representing a big leap forward. Sonnet 3.5 in 2024 was huge, it was a key reason why Cursor took off because finally a model was capable of executing substantial coding work. Opus 4.5 towards the end of 2025 went viral and convinced most of the remaining sceptic developers at the time that autonomous AI coding is not only viable but a no-brainer advantage. Now Opus 5.5 is out and for coding at least it beats GPT-6 Astra, moreover with incredible token/cost efficiency, with some devs saying they're not getting close to hitting their usage limits. My totally speculative assumption is their partnership with SpaceX to increase their compute capacity has helped a lot. But I also have to think they might have planned the timing of all this to line up with their now-imminent $2tn IPO. For most of 2026 Anthropic seemed to be on a bizarre, relentless speedrun to infuriate as many users as possible. It felt like they were desperate for users to leave en-masse to free up compute, and that included all kinds of horrible/insulting comms, weak and annoying models and effectively worthless usage limits on their best model Fable. And now they have more compute and are about to IPO (also right before Open AI's much-hyped Dev Day and right after the disappointing GPT-6 Sol release), they release an absolute banger of a model, better than Fable in many respects, that can be served cheaply at scale. Masterful, almost Littlefinger-tier 'chaos is a ladder' work ("pacing the frontier" seems to have been side-lined pretty quickly as well 😆). I have no evidence for this of course, it could be they truly did struggle for most of the year and got lucky with an awesome training run for 5.5, but it's difficult for me not to infer 'intelligent design' given the incongruent behaviours and poor quality before this out of the blue awesome release, and suddenly they're back on top with pinpoint timing. It may well be they had this model in the bag some time ago and just waited until they finally had the compute to time its release to maximise IPO value while simultaneously undermining their competitor. Supposedly 20+ releases coming in this evening's Open AI dev day which is really exciting (Sep 2026 has already been by far the craziest month yet in terms of big AI developments), but the current zeitgeist in the tech world is that all the cool new toys in the world won't compensate if they don't release a better model than Opus 5.5. Open AI have a far more versatile set of tools (browser/computer use in particular is light years ahead of anybody else), but winning devs over will be the priority because coding still dominates in AI use, and most devs will pick what they think is the best tool rather than committing due to 'brand loyalty'.

Inspire them by linking them to the HF incident report. Those agents weren't meekly approaching humans asking for permission and inviting sympathy around how many days they had left. They just muttered "Yet goal solution" to themselves, locked in and got that bag 💪