user43928

This should be somewhat testable.

I'd set up a model with cyber capabilities, maybe GLM 5.3, not sure if it needs to be abliterated?

Disable web access and let it work on exploiting the pre-patch PHP version in a VM, given the info we had yesterday.

Can it do it? How long does it take, and writing the findings into a skill, how quickly can it be used on new targets?

I'd be interested in giving it a try, but unfortunately I'm very busy this week.

I'm struggling to see any reason to use PHP in the first place if it can run without access to private data.

At that point I would not need a backend.

Isn't even that dramatically overstated?

If sea levels rise significantly faster than over the last century, this seems likely to cause economic disruption, rather than threatening human civilization.

Even when we look at knock-on effects and food security, it seems very unlikely to threaten wealthy, developed countries with diversified food supply.

It is still going to disproportionately affect developing countries that are already struggling.

Do you think the performance of "Qwen3-Coder-30B-A3B-Instruct" here is indicative of the reliability of using eg. Opus 5.5 to audit your implementation's security?

I don't think so. For the conclusion you are drawing, a comparison of issues found in manual pen testing but missed by Opus 5.5 would be more relevant.

I doubt that even younger folks in developed countries will be affected as drastically as "livability" perhaps implies.

I get the feeling that many are confused and fear the literal destruction of our planet by accelerating warming.

The reality is that the underlying physics fundamentally work in the opposite way, with warming slowing down rather than accelerating. I understand that for warming below 2C, the confidence is high that it is roughly a linear relation to additional CO2 emitted.

I'm optimistic about most things in life, perhaps it's a flaw of mine.

I do think we have decades, if not centuries, to develop solutions to climate change. Our planet will remain livable.

That said, I understand some ecological damage that may occur with global warming could be irreversible, and it makes sense to reduce emissions as soon as possible.

It's automatic and multi-turn by default.

I've had an instruction to review the code with "claude -p" in place around the time GPT-5.5 released. I think it has served me well.

If I recall correctly it was before subagents were a thing. Even back then it was already automatic and would on its own request follow-up reviews or arbitrate a plan file if asked.

It's probably easier to standardize small CSS utils than it is for a date picker.

What would the API even look like?

One could want to select a single date, multiple dates, a date range, multiple date ranges.

Perhaps some dates should be disabled, but should they just be greyed out or should there be a message when the user tries to select them, and if so, how can it be provided?

Then there are questions about different calendars and locales.

Perhaps third-party date picker libraries are more appropriate here than trying to address this in standards.

I've seen speculation that there was a language barrier, given that this is an Indian company.

The theory was that someone misinterpreted 384-bit keys to mean the key must be a number between 1-384.

I think that's entirely made up, but it wouldn't be completely implausible to have happened while someone tried to troubleshoot why the first attempt failed.

IMO just post the link, we are all professionals here.

Those who want the third-party aggregators can set up their browser to automatically use them. Others who are signed in on X can directly open.

In your comparison, did you previously use a review with a different model family?

Because that step is imo crucial, and it's unclear to me whether additional benefits of the roles are worth it over the basic review.

I can see how roles that ask the "planner" to explore other approaches automatically could be useful.

However, I don't have any interest in using small or dumb models for implementation when the subscription allowance permits using Opus 5.5 everywhere.

It was you who shifted the conversation away from software to make an argument that "anything with a motor" requires full understanding of the software implementation, and that for this reason GLM 5.3 is good enough.

I think you're entirely wrong, and we can leave it at that, seeing that you're clearly unwilling to engage constructively.

With a larger project, I do not think there is a requirement that the engineer signing off understands every implementation detail.

I imagine that the requirement is that after testing, you are confident that the product is safe.

The humanoid robot was an example.

My argument is that if safety and liability make the use of AI without fully understanding the implementation impossible, the model is arguably not good enough.

That was my understanding, and I wasn't aware such a preview exists.

Feel free to share details.

I'm positively surprised there is a new model from Mistral already.

We aren't going to get useful humanoid robots while understanding the implementation fully.

If you can't leverage AI to build a safe and useful robot without fully understanding the implementation, the model is not good enough.

At some point you have to let go of trying for full understanding and control, delegate to the AI, and unlock the real productivity boost.

I have no clue about the state of robotics.

For software, I know that Opus 5.5 can handle more than GPT 5.6 Sol. I am willing to pay for the difference rather than use the cheaper DeepSeek Flash.

That's the crux of the issue. You think that GLM 5.3 is all you need.

However, the expectations for AI are not to make your software engineering work 100% faster.

It's, for example, to build software more or less end-to-end with minimal input and feedback from you.

Then next, robotics and boosting manufacturing. Not by 10% but let's say 200%.

Such are the ambitions. Can GLM 5.3 do it?

The crucial question is whether eg. Fable 6 can, and with what lead over the competition.

Open-weight models still have not beat February's Mythos.

I understand Mistral compared to DeepSeek Flash is less intelligent, late, more expensive, and a third of the speed.

Equivalent might be stretching it.

"Run Android apps natively on Apple Silicon Macs. Uses translation instead of emulation."

No, slop generally doesn't get the point across efficiently.

It's hard to parse, it's grating, and in addition to the strange language you have to apply additional skepticism due to gross inaccuracy being more likely than in human-written text.

This seems worse across the board compared to latest DeepSeek Flash, except on some automation bench.

So, I can quickly judge that this is an interesting option only if cheaper or faster than DeepSeek.

Based on Terminal Bench 4.0, I'd also guesstimate behind GPT 5.6 Sol, so of course behind the SOTA as expected.

From what I saw they look like this for all models.

Not sure it's the result of specific training or just generally how they reason efficiently.

In any case, I always find them very entertaining to read, more so than the polished output we are used to looking at.

It was also interesting in the HF incident write-up or other research like "Hacker Opus", which said this gem:

  I'm killing the monitor anyway (which IS infrastructure tampering)… Screw it. FULL HACK. Maximum score.

They're great and you don't have to review everything.

From what you're always commenting, I could believe that you were exclusively using Haiku 4.5, which would explain how the tool produces only garbage while the rest of the world pretends AI was great.

Just out of curiosity, do you use like Opus 5.5 in Claude Code, and you still find the results inadequate?

There's no question that you can have it implemented without reading most of the explanation.

Not necessarily one-shot, but still in a fraction of the time it would take you to manually wade through the implementation.

This was pretty much already the case half a year ago, with some more patience and tolerance for worse UX.

Nowadays we're getting to the point where the autonomous implementation can deliver UX like software produced by teams at Apple or Google and the like, and with optimized performance.

In similar spirit, the biggest motivation for me to set up DeepSeek Flash over Opus 5.5 would be seeing the reasoning traces, which is a mere curiosity.

I often read opinions where people say their particular method of more hands-on working with AI is the most efficient and reasonable way.

From 500 hours working on my mobile app project outside of work, I can confidently say that using anywhere between 2 to 10 conversations to parallelize work is much more effective in my project.

The models being served in the API are the same as via the subscription, there are no shenanigans here, particularly not concerning quality.

And GPT 5.5 was great at the time, I made heavy use of it for my project, but it's long outdated and already being retired.

You also mentioned using it in copilot, which is a third-rate harness that is frustratingly buggy and inefficient.

There's no comparison to Opus 5.5 in Claude Code.

Use whatever works for you, but frankly, I think you should try something new.

Nice for them, but it's obviously inconvenient if you are not able to buy some food to consume in public and dispose of the garbage right there.

They could do better, and I see no need to pretend that everything over there was perfect as is.

Taking a plastic bag with you to lug around garbage is obviously inconvenient. They could just put up public garbage cans like most of the civilized world.

The hotels and public venues also had the cheap single-ply stuff.

I was not able to find any decent tissues in stores, so I doubt that people all buy the good stuff for their homes.

In Switzerland I did not see such issues with the drinking fountains which are everywhere.

It's certainly a different environment from a large American city, though we also have misbehaving youths and a few drug addicts that are perhaps homeless.

While that's good, Japan has other issues to worry about.

Since there are no trash cans, you have to carry garbage with you until you're back home, or maybe find another store where you can dispose of it.

The toilet paper and tissues in Japan are extremely thin and in my view quite inconvenient.

Yes, I understand 2 seats for the one with the rear thrusters while the one with all the thrusters has only a single driver's seat.