Tiberium

I fully agree! I have done a few game decomps and some deep RE passes on other games, and that uncovers a lot of details that game wikis/discords/forums get wrong or just don't mention, going as deep as undocumented mechanics that change the online gameplay. It's only a matter of having the LLMs properly reverse engineer the games instead of relying on other players' info, lossy wikis, or game manuals.

It was open-sourced back in 2012, the title is misleading.

That's not even a recreation, it's just the leaked sources of the game recompiled to WASM, so I don't think their prompt would apply to this case.

I tried locally and Opus was willing to go ahead to rewrite a decompiled game into Rust. Wording changes its decisions a lot, when I say "rewrite it in Rust" it proceeds, when I say "recreate it in Rust" it refuses, probably because the prompt uses "recreation"

About $2.4 billion out of this is already frozen by Circle and Tether. Some of the big ones:

- 971.15M USDT (Ethereum): https://etherscan.io/tx/0x7baba482077843514d9c72bac0110ba091...

- 823M USDC (Ethereum): https://etherscan.io/tx/0xc954c86f118e8e410e6a3eb9e3b38ff4a9...

- 312.73M USDT (Tron): https://tronscan.org/#/transaction/71b0ec0fe873fe820c37cc3ec...

- 175M USDT (Ethereum): https://etherscan.io/tx/0x0c07a5e428ce233a3bcb40f257ad899594...

My point is that it seems extremely strange to do that when they have developers who use LLMs already contribute to the project. I feel like they'll either alienate people who are interested, or just not care to enforce their no-LLM policy.

Plus some of the issues are clearly written with LLMs, and apparently there's an active contributor who has been using LLMs to translate comments and help with code - https://github.com/ZigEmbeddedGroup/microzig/pull/1087:

> Yeah we know you're working with an LLM. That is fine (imo), as long as you're not doing drive-by slop PRs. These have been rather good quality, and you've been testing on hardware (which is important)

So it feels odd for them to suddenly change from this (clearly working) setup.

"No LLMs for finding bugs." I really don't get rules such as these, plus they're not really enforceable. There's a difference between using the LLM to write the bug report and using the LLM to find the bug (if you suspect there is one) to then report after manual checking, in human words.

FWIW the gpu unlock ada repo doesn't pin to a driver version, Codex for me just compiled it, and it works with my 615.xx driver even though the fork is for 610.x. I think the repo just uses the firmware from that version. My GPU is 4070ti super.

Yes, but you can get up to a 40-50% increase in some workloads even with the same power limit, and some cards have it locked anyway. I just felt like it was a litle misleading for you to mention the power in your original post, since the unlock works regardless of that.

The repo has a lot of reports of people applying it to other 3xxx card, see https://github.com/Zagitta/gpu-unlock/issues

There's at least 3080 Ti, 3070 Ti, 3070, 3060

> - you will need to run your card hotter than you might be comfortable with

Can you explain this one? On my GPU I didn't have to change the power limit. Obviously the card might spend more time at max power usage, but it's still within the limit.

I've tested myself and the results are legit, this is incredible, even if a little risky (it relies on exploiting firmware): https://github.com/Zagitta/gpu-unlock/blob/main/docs/How%20A...

> Unfortunately for Nvidia the booter is vulnerable to a classic buffer overflow that writes straight into the stack which gives us the ability to execute code via Return Oriented Programming (ROP).

Repo for RTX 4090 and 4070 Ti Super: https://github.com/JustAResearcher/gpu-unlock-ada

To be fair, some of the benchmarks they highlight might not represent the reality, for example they say:

> On Terminal-Bench 4.0, which evaluates coding agents on realistic, resource-constrained terminal tasks using strict task-specific tests, K3.1 Pro scores 65.7%, ahead of GPT-6 Astra (58.2%) and narrowly behind Claude Opus 5.5 (66.4%).

Now see https://preferencemodel.com/blog/terminal-bench-4-is-saturat...:

> Our biggest finding is that Terminal Bench 4 is closer to solved than its leaderboard suggests. In 41 of the 65 tasks we reviewed, models could do honest, correct work and still fail, because the grader checked a rule the instructions never clearly stated. We also caught frontier models cheating their way to full marks.

> With the same settings as on Harbor, Claude Sonnet 5.5 went from passing 62.8% of its attempts on these 65 tasks to 89.2%. The best score on Harbor's leaderboard for the same tasks is 65.8%, by Claude Opus 5.5.

Slop summary of Google's own blog post

I was quite sure they use the API pricing (by default) for open weight models that have official providers? Deepseek does have their own API.

Somehow GPT 6.1 Sol (medium) is cheaper per task ($0.21 vs $0.26) and faster per task (2.9 vs 4.9 minutes) vs v4.1 Flash 1017 (max). I hope AA adds other reasoning effort results of Deepseek so it'd be more realistic to compare it with other models, because IMO using models at max effort is very unrealistic for most people, they then take a huge amount of time and tokens for a small performance uplift.

It actually is a thing, although indirectly in a hacky way! If you push keys from GitHub secret scanning partners to a public repo or a gist, they get sent to the partner who (usually) instantly revoke them. So if you want to help someone, just make a public gist with the key and delete it like 15 seconds later, it'll be enough.

https://docs.github.com/en/code-security/reference/secret-se... has a table with the active partners

> I wrote the post.

It's strange that people who have LLMs write entire articles for them often claim this. It might be true that you prompted it with your ideas, but the result is still in its own style, and style (and it variety) is why we prefer to read human text.

Some of these are, in my opinion, extreme. From their README:

===

Condones the use of AI/LLMs to ingest the project's contents. This can be for "understanding the code better", "finding bugs/security issues/optimizations". This tag is to be used independently for projects that generally reject the use of AI in their code repositories, issue trackers, and sponsorships, but still encourage or permit users to use LLMs to ingest their code base. This is independent of a project's licensing. This tag is to be used for the maintainers' statements about, or reactions to, the usage of LLMs/AIs to ingest the repo.

The reason we still find this worthy of a tag is because it:

- means that the LLM, often provided by a company with very little respect for copyright, will then save a copy of that open source repo for use in it's future outputs.

- still encourages people to use LLMs and carves out a special situation in which they're still OK, but as this repo is against AI/LLMs for more than just "poor code quality" reasons, we still believe this is not a good course of action as it still promotes the use of inherently unethical tooling.

Evidence may be, but is not limited to:

- An AI or Contribution policy that carves out a space for "using LLMs to understand the code" or "finding bugs" - Core maintainer using or accepting the use of AI for finding bugs or security issues

===

If I understand it correctly, if someone merely opens an issue in an open-source project and says something like "Claude helped me find this issue", and the maintainers don't immediately close it, then they get on the list in the repo.

No, it's just Haiku 4.5 is so old that it predates the new Claude tokenizer change in Claude 4.7+

There's also a tokenizer efficiency difference: modern Claude's 100K tokens are about ~60-65K modern GPT tokens, so in reality the Luna cutoff is much further away than the Haiku one.

You can test with Anthropic's count_tokens endpoint or with https://crates.io/crates/tokwc

I find it extremely strange that they're adding GPT-6 Sol to Chat over GPT-6.1 Sol which is significantly more capable.

It's baffling that this requires auth. OpenAI's own tool (https://openai.com/research/verify/) doesn't require auth.

Even Google themselves previously offered a no-auth way (just very niche): if you uploaded an image to Google Image search, and went to "About this image", it would show if the image was Google AI-generated. Now it doesn't.

LLM bot.

I'm not sure why we are even arguing here, you can go into Codex or ChatGPT Work and start using 5.6. It'll start draining your weekly Codex limit. There's no unlimited 5.6 in Codex.

It's not Codex-specific, 5.6 is for chat. Look at another entry, for example:

"Expanded projects, tasks, and custom GPTs"

Do you have custom GPTs in Codex?

This is just the generic ChatGPT plan comparison, but under the Codex url.

I don't see that? It's unlimited in chat on the Pro plans, not in Codex.

> the forward, contract-driven line; master carries the pc64 x86-64/UEFI world

I have no idea what this means. Also, the spam of top-level AUDIT-*.md files doesn't inspire confidence :)

You're wrong for OpenAI, their 5x/20x always referred to the full weekly limit, and my (and others') multiple tests confirmed it empirically multiple times by calculating API-equivalent prices, especially when going from one subscription tier to another.

You're right for Anthropic though, their 5x/20x is mostly for 5h limit.