novaleaf
novaleaf on Hacker News
I think nobody threatens NK because they have 10k artillery pieces in range of SK capital, and have all sorts of NCB munitions to use.
can you recommend any benchmark websites that show up-to-date details like this?
TerminaBench, DeepSwe sites are out of date.
Another comment approximates at around USD$15k, so yeah, not zero.
> "Adversarial red-teaming made auto mode stronger"
Has to be said: they (and OpenAi) don't let you red-team your own creations, pisses me off to no end. </rant>
opus 5 doesn't follow instructions that great. probably needed that extra "creativity" to benchmax. if you are stuck on anthropic, try Opus 4.8 or Fable5 a try with the same prompts. very different results.
that's pretty nice illustration work, I'd agree, keep doing it :D
I ranted about this in a prior thread [1] but just now, when I ask Claude to verify my (existing) app's protection DRM system works, it terminates my request:
> `API Error: Connection closed mid-response. The response above may be incomplete`
Claude doesn't have a "Security whitelist" for small biz. Codex does, but they never replied to my application. This is a great example why, as of today, everyone NEEDS access to the Open Weight models. Frontier models can hack you, we should have access to tools assisting defense.
[1]: https://news.ycombinator.com/item?id=49035303#49040674
as others mentioned, it sounds like "Off to Be the Wizard". Unfortunately I wouldn't call it a great series. It starts off strong but drags. IIRC I got through half of the second book before stopping.
With Openai and Anthropic, you can opt-out of training. Kimi you can't.
> I guess part of it is also that I don't mind doing 'hand-edits' like for example LLMs love to say "// so and so removed" I just go and remove that manually later rather than being like "don't comment about what you removed!11" cause you're really fighting deep grooves in the model's behavior at that point.
Maybe I'm misunderstanding you, but that's just about the best example possible for using AGENTS/CLAUDE.md. Just add "don't comment about what you removed!11" and you never have to say it again...
...but you'll get constantly nagged about the `11` of course!
Agree, out of the two, I can get Codex to design and implement security systems (Fable just refuses to discuss). I've only used K3 a little bit, but I found that being able to discuss attack vectors and their mechanics gives me details to paste back into Codex for it to ingest. I could never have gotten there through Codex alone.
Yeah it's not cheap, it's just the normal $200 subs for each.
I applied for individuals. I did their verification steps and answered questions in a few minutes, so that was indeed easy. The problem is that was all that happened. No followup, no access, no denial. When I tried to reapply it tells me I can't apply again.
OpenAi (and Anthropic) have no incentive to allow security access to individuals. I'm not a deep-pocket org or influential gov agency. Allowing individuals increases the risk of bad press (what if I do something naughty and talk about it?) so best to ignore us.
keep in mind that Kimi trains on your data, so if you care about that, sandbox what it has access to.
I think security is an integral part of any production software, and if you get value from LLM's in software development, it seems likely they can be useful in security too.
My point and frustration is that gatekeeping in the name of Security makes the Chinese models actually better at security than USA models.
I don't think it's nefarious, but the end result leads to a pretty frustrating experience by anybody needing actual security work. (and without the organizational deep pockets to obtain SOC 2 attestation)
Some irony: I subscribe to Claude and Codex (20x plans), and now Kimi.
Why Kimi? because K3 is the only frontier model I can have a serious conversation with about my product's security.
(I did apply for OpenAi's Cyber Pilot but got no response)
Reminds me of the gain-of-function, COVID lab leak hypothesis. It seems like humanity just can't stay away from Pandora's box.
frontier vs "not quite" :D
makes me wonder if the starbucks story is fiction too :P
no, I didn't get access to sol until a few hours ago. I just have my claude protocol files linked to inside codex. trying Sol this morning for the first time, so I can't really comment on that vs gpt5.5.
However you can do what you are asking "fable--> sol" you need to setup a mcp or have fable run a bash tool, just invoke the `codex.exe` cli tool with whatever cmdline args are needed.
I sub both codex and claude at 20x. I like opus+fable more than gpt5.5 because it seems gpt tries to finish tasks by leaving any ambiguity unresolved. claude seems better at surfacing open questions.
This is using the same AGENTS.md prompts, which were designed firstly for Claude use, so maybe it's something that could be optimized better if I understood gpt as well?
I think the story was Starbucks -> Seattle's Best Coffee, not McD's --> BK. but it does work I guess.
reminds me of SBC's (Seattle's Best Coffee) strategy, which was decidedly not Nash: put a store across the street from every Starbucks.
it works by ppl harvesting low ranked submissions and resubmitting them via a (unofficial) voting ring.
I used to live in Thailand, and over the last 10 or so years it seems that the kind of DIY content you find on Youtube has had a positive effect. Before then it was difficult to find a building maintenance team that understood the purpose of a waste water P-Trap (unfortunately I'm not exaggerating)
yeah, for example, just send a hash of the domain used. but then maybe people would say anthropic is spying on everyone, instead of targeted spying...
it was released a few hrs ago as "Fable 5". it's an incremental improvement over Opus 4.8.