user43928

It would be a low quality training set that you theorize is a goldmine and worth illegally stealing from your customers.

I think this data is likely worthless compared to curated RL tasks.

Good luck to the accountant that tries to tell leadership to slash AI usage.

I'm sure investors will love it.

Now we're starting to see real impact from AI, people are learning how to use it, and OpenAI cut prices by no less than 60% like a week ago.

You think now is the time they're going to cut the spend?

If you assume that noisy general usage data enables good RL, particularly compared to curated RL training sets.

I am not convinced that's the case.

Accurate training data?

At best you produce some noisy signals that are going to have a tiny impact if even that.

And that's on a personal plan where you didn't opt out of sharing usage data.

Business plans offer zero data retention. This is a non-issue.

Presumably everyone who doesn't bother integrating a third party API key into their harness, which would probably be most of the Claude Code users.

I never noticed downtime with public GitLab. It's also generally better than GitHub and Actions.

A self-hosted enterprise GitLab instance in Azure, on the other hand, has had outages and performance degradation quite a few times.

I wonder how relevant this really is.

How hard can it be to download a set of real images and generated ones in order to train a model to detect and strip the watermark with minimal perceptual difference?

I'm just wondering if it's really the result of impressive innovation, in line with comments saying that Mistral might in the future beat a bankrupt OpenAI due to being more profitable or efficient.

Or if it's more a situation where I could have Opus 5.5 go through recent papers from the Chinese labs to apply these to a new large training run, resulting in a model that performs similarly to Mistral Large 4.

I honestly don't know.

Not our field? Most of us are programmers here.

It is clear to me that any competent person with a little patience can now build software better than what I used to build by hand.

It's better than I expected for Mistral.

That said, it hardly seems like the innovative underdog some comments here make it out to be.

GLM-5.3 and Kimi K3 released 2 and 3 months ago respectively.

For all I know anyone could train a model of similar performance by just applying already published research to a new large training run.

> The wholesale contempt that so many AI users have for human coding demonstrates that it's an entirely different culture with different ethics and different goals

Personally I enjoyed building apps before AI, and have been doing so professionally for 10 years.

With AI it's much more fun for me since productivity is way up.

It's not that I have contempt for human coding, I just know from my own experience and hard work that using AI is much more productive.

People online constantly make comments where they insult my expertise and the quality of what I am building. Surely you can see how this could quickly lead to a heated discussion.

I don't think that's correct.

> Following an investigation, we have confirmed that Buckmaster’s Codex prompts over the two months preceding this announcement and paper on September 8, 2026, could not have influenced the system in any way, including through training. The OpenAI internal model used for this result was developed through large-scale reinforcement learning on top of a previously pretrained model. Our proofs also differ significantly. In the Euler case, Alpöge and Buckmaster proved a result with external forcing, while OpenAI’s system proved a result without external forcing.

Doesn't that seem exaggerated when clearly people are gaining the freedom to modify any software easily and quickly?

Last week, I've seen videos of people merging video games like Super Mario and Minecraft to create absurd new games that previously you wouldn't imagine anyone implementing because the cost/effort was prohibitive.

My app is relatively self-contained.

Complexity is up for debate. The project currently has ~1500 en translations and 7 AI models I optimized for mobile.

I think the key point is being self-contained, so that QA testing is effective for catching issues.

These comments were also around for the Opus 4.5 release.

Prices for the same task (ARC-AGI-2) have since dropped by factor 10 if you compare to Opus 5.5 Low.

I'm wondering, do you check out MRs from forks in your IDE? What tooling do you use?

On the command line it seems a bit tedious with forks, and disk space constraints in our crappy enterprise environment are another factor.

That's a pretty high bar.

By that standard, couldn't I say you don't truly own open source software like Chromium or OpenSSL because very few are capable of understanding the codebase well enough to implement changes?

Many on HN still have the opinion that you must understand every line of code in the project, and that all is lost should you merge code that wasn't reviewed.

Obviously any model will do if you use it as a better autocomplete.

I believe that there is a large gap in expectations between different workflows.

Until the AI like reads my mind and produces perfectly production ready apps with minimal intervention from my side, there is still going to be room for improvement.

There surely is code so critical that not fully understanding every line of it is not an option.

That's probably the rare exception.

Even in finance everyone is using AI.

You'd be wrong, I'm working on a paid mobile app, with the express goal of delivering a better product than existing competitors.

I've put like 500 hours of work into it over the last 6 months, besides my full-time software engineering job.

Frankly I think I have more experience than most here at using AI.

For me it works very well to just have a conversation with the agent in order to understand issues and steer it.

Essentially I would ask the question I would otherwise try to answer by reading the code.

Obviously, this has the downside that sometimes you miss things that would be obvious when reading and understanding all the code.

Since that can take a lot of time, it is still more efficient for me to instead QA test functionality and address issues I notice during actual use.

I do get that some people enjoy the manual process more than they enjoy building the product.

It doesn't give you the right to insult my expertise, and frankly your arguments sound like things that one would make up despite little actual experience using AI to develop software.

I always preferred to check MR diffs in GitLab.

And I don't use the debugger nowadays. The AI can do it much faster and add relevant instrumentation and parse logs.

Only that it isn't, and that millions of professionals, including myself, use agentic AI successfully to implement high quality software.

That there is no lack of people delivering low quality using AI does not change this.

Using a CLI coding agent in your IDE's terminal is a terrible way to view conversation history and works poorly for organizing multiple conversations.

It only fits a hands-on, merely AI-assisted development approach, on one feature at a time, that is not typically productive compared to agentic development.

You might disagree, but you'll find that the proponents of the hands-on approach are losing ground every month as AI gets better at working autonomously.

Not sure who downvoted this. JetBrains IDEs are done for.

I don't even remember the last time I opened them, I cancelled my All Products pack this year.

The advantage it had over vscode was always the better language features and the familiar UI.

None of that is useful to me in AI based workflows. Some might still disagree, but the manual development workflow using an IDE is clearly dying.

JetBrain's revenue will obviously crater as a result, particularly when slower enterprise licenses follow.

It should be obvious if current AI has been used to write any significant part on its own.

I don't consider investing $20k into hardware to run SOTA open models, that are far behind proprietary SOTA, to be competitive.

Even if open models were competitive, it's still typically going to be more expensive than a cloud provider because of low utilization and higher purchase price.

I don't think anyone criticized the timing of the patch.

But I find it egregious that they didn't roll it out as a security update at all, which is why it was not automatically installed in OP's case, even though the fix was already available.

I mean, what else requires a hotfix via security update if not a fatal flaw in your remote access authentication leading to full root access, that is actively being exploited in the wild?

Also, it's not really on the user to gate remote access behind an additional firewall and authentication layer. This is something that just has to work securely.

If it doesn't, that's understandable, but still hardly the user's fault.

Where does negligence come into play?

In criminal hacking law that explicitly requires intent, or a civil lawsuit about damages?

I honestly don't know, but my guess would be the latter.

I read earlier that the crime is making such threats in any manner where another person could see them.

I do not know whether it matters if the suspect knew that their chats could be viewed by Anthropic staff, or if the only thing that matters is whether another person was able to view the chats.

Edit: Upon searching I now also agree, it seems to require a state of mind and perhaps awareness of the possibility that the threat would be seen.