rectang

(Not that there aren't elements within the Trump administration who might contemplate a false flag operation, but it will be obvious if they try it because they aren't clever enough to pull it off.)

Furthermore, both Ceaucescu and Mussolini were extrajudicial executions.

Nixon is a better precedent for the USA because it demonstrates that we're already capable of ejecting a criminal president without bloodshed. It wasn't a perfect process (the Ford pardon was terrible) but in the end Nixon still lost power.

Will a future administration have an opportunity to build something new and better from scratch which would not have been possible due to institutional resistance before it was all burnt down?

I anticipate a Napster-style reckoning at some point when there's a successful high-profile copyright suit around obviously derivative output. It will probably happen in video or imagery first.

I'm a BSD-style Open Source advocate who has published a lot of Apache-licensed code. I have never accepted that AI companies can just come in and train their models on that code without preserving my license, just allowing their users to claim copyright on generated output and take it proprietary or do whatever.

I would actually not mind licensing my work in an LLM-friendly way, contributing towards a public pool from which generated output would remain in that pool. Perhaps there is opportunity for Open Source organizations to evolve licenses to facilitate such usage.

For what it's worth, I would be happy to pay for a commercial LLM trained on public domain or other properly licensed works whose output is legitimately public domain.

The market-leading technology is pretty close to "good enough" for how I'm using it. I look forward to the day when LLM-assisted coding is commoditized. I could really go for an open source model based on properly licensed code.

I feel like if we're going to make progress in preventing wholesale data breaches it will be through architectural innovations that attack the problem of why a trove of concentrated data needs to exist. Even if the government needs to be a central authority, are there ways to house the data that limit the blast radius?

I'm sure there are innumerable arguments why this can't help, but when the mainstream alternative is despair and helplessness, progress will be made in the margins.

From what I can tell Opus 4.7 is more resource-intensive than Opus 4.6 is more resource-intensive than Opus 4.5.

I'm just choosing to focus on the substance of the argument itself, which I think is risible regardless of who makes it and why.

From that page:

> Today, AI can be pointed at an open source codebase and systematically scan it for vulnerabilities.

Yeah, and AI can also be pointed at closed source as soon as that source leaks. The threat has increased for both open and closed source in roughly the same amount.

In fact, open source benefits from white hat scanning for vulnerabilities, while closed source does not. So when there's a vuln in open source, there will likely be a shorter window between when it is known by attackers and when authors are alerted.

As far as I can tell, the distinctive feature of my workflow is that I'm giving it small, contained single-commit-sized tasks and limited context. For instance: "For all controller `output()` functions under `Controller/Edit/` and `Controller/Report/`, ensure that they check `Auth::userCanManage`." Others seem to be taking bigger swings.

Now it's going to cost me an upgrade to $39 Github Pro+ to keep using Opus, and even then it's with much higher multipliers. I don't fully understand the extent to which this reflects actual costs for Opus versus Microsoft leveraging network effects to discourage the usage of a competitor.

I didn't really want to wander outside of VSCode just yet because I was happy with VSCode/Copilot/Opus-4.5 and I don't want to spend all my time experimenting when stuff is changing so fast. But I guess my hand has been forced.

I'm closely supervising the LLM, giving it fine-grained instructions — I generally understand the full interface design and most times the whole implementation (though sometimes I skim). When I have the LLM write unit tests for me, it writes essentially what I would have written a couple years ago, except that it tends to be more thorough and add a few more tests I wouldn't have had the patience to write. That saves me quite a bit of time, and the LLM-generated unit tests are probably somewhat better than what I would have written myself.

I won't say that I never see brain-dead mistakes of the "5-vertex square" variety (haha) — by their nature, LLMs tend towards consistency rather than understanding after all. But I've been using Claude Opus exclusively for while and it doesn't tend to make those mistakes nearly as often as I used to see with lower-powered LLMs.

??

You say you deleted the tests, because you "should test it"? The logic seems inconsistent.

Sanity checking LLM-generated code with LLM-generated automated tests is low-cost and high-yield because LLMs are really good at writing tests.

It seems like they're working hard to prioritize wrapping their arms around huge contexts, as opposed to handling small tasks with precision. I prefer to limit the context and the scope of the task and focus on trying to get everything right in incremental steps.

I find that Opus is really good at discerning what I mean, even when I don't state it very clearly. Sonnet often doesn't quite get where I'm going and it sometimes builds things that don't make sense. Sonnet also occasionally makes outright mistakes, like not catching every location that needs to be changed; Opus makes nearly every code change flawlessly, as if it's thinking through "what could go wrong" like a good engineer would.

Sonnet is still better than older and/or less-capable models like GPT 4.1, Raptor mini (Preview), or GPT-5 mini, which all fail in the same way as Sonnet but more dramatically... but Opus is much better than Sonnet.

Recent full-powered GPTs (including the Codex variants) are competitive with Opus 4.6, but Opus 4.5 in particular is best in class for my workflow. I speculate that Opus 4.5 dedicates the most cycles out of all models to checking its work and ensuring correctness — as opposed to reaching for the skies to chase ambitious, highly complex coding tasks.

> Over the coming weeks, Opus 4.7 will replace Opus 4.5 and Opus 4.6 in the model picker for Copilot Pro+[...]

> This model is launching with a 7.5× premium request multiplier as part of promotional pricing until April 30th.

This gives me hope that even if future versions of Opus continue to target long-running tasks and get more and more expensive while being less-and-less appropriate for my style, that a competitor can build a model akin to Opus 4.5 which is suitable for my workflow, optimizing for other factors like cost.

My workflow is to give the agent pretty fine-grained instructions, and I'm always fighting agents that insist on doing too much. Opus 4.5 is the best out of all agents I've tried at following the guidance to do only-what-is-needed-and-no-more.

Opus 4.6 takes longer, overthinks things and changes too much; the high-powered GPTs are similarly flawed. Other models such as Sonnet aren't nearly as good at discerning my intentions from less-than-perfectly-crafted prompts as Opus.

Eventually, I quit experimenting and just started using Opus 4.5 exclusively knowing this would all be different in a few months anyway. Opus cost more, but the value was there.

But now I see that 4.7 is going to replace both 4.5 and 4.6 in VSCode Copilot, and with a 7.5x modifier. Based on the description, this is going to be a price hike for slower performance — and if the 4.5 to 4.6 change is any guide, more overthinking targeted at long-running tasks, rather than fine-grained. For me, that seems like a step backwards.

With pro analog tape recordings (e.g. 2-inch 24 track, half-inch 2-track), you record alignment tones onto the tapes to capture the state of the recording device, and then later calibrate the playback device to the particular tape so that playback alignment matches recording alignment. But this is essentially never done with cassettes, so you have to earball it.

Cassette players for mastering studios actually have alignment options (e.g. adjustable azimuth) that aren't present on consumer devices. But without the tones, you have to guess.

The problem with starting from a digitized source is that it may have been digitized from non-aligned playback. Ideally you want to go back to the analog originals - but old cassettes are rarely in perfect condition.

Elsethread, you mention that "some people really got upset.". In some sense, the more upset they get, the more successful the troll and the funnier it gets, right? At least, it feels funny to me, at the same time as it also feels bad to imagine upset kids, at the same time as feeling that upset kids learning that other humans are cruel is a necessary part of growing up.

Pricing mistakes which make the supermarket money are unfortunate but low priority. Pricing mistakes which cost the supermarket money must be fixed immediately.