From what pros are saying after matches, especially in classical chess, it's almost always preparation and metagaming the opponent vs finding good moves.
It's a lot of mind games involved, going for specific openings, repeating same moves as last match, doing a poor by surprising, move etc.
Many GMs said that to win a match you have to play non-ideal moves many times, otherwise it will likely end in a draw.
Well people will still pay to watch a human player play tennis even if a robot could play infinitely better than that human. Try something like that with an employer and software engineer combo (and no, you don't even have to imagine that).
Well look, the wall is already losing every point by volleying your serve so I feel like I'm on pretty strong ground here already. But give me an exact specification for your curved wall and I'll tell you how I'd beat it.
It seems that a rogue engineer poisoned the prompt in that instance. But the fact that they keep the system prompt open is nice. Generally I am biased towards favoring more freedom and openness rather than clamping it down in the name of safety.
idalib with Claude Code already works really well. But honestly, despite what people have been saying, LLMs have been very good at decompiling for at least two years now, I have been using it for that purpose regularly. Even just copying disassembly from the current function and all nested called functions from IDA into ChatGPT is already unexpectedly good.
The thoughts you are having about this show me that you are at a local maximum, but you will find your way out if you keep trying everything. There is no single thing that will help. Read novels, copy them word by word, translate them, write short stories, dissect stories, stop movies halfway and ask yourself how you would continue them. Read a ton of writing advice, try them out but don’t take it as gospel.
Even if you don’t improve for 100 days straight, small successes accumulate. In a decade you will have transformed yourself.
There is just no way you won’t improve significantly if you keep trying new things and bring yourself to fail ever day.
What helped me was the saying “your first million words are gonna be shit”. I still distinctly remember, four or five years into writing every day, when things finally clicked, my voice came through, and my sentences became fun. It is delayed gratification to the max.
I must say I expected an actual poisoning of the data used to train the LLM and was excited, but the examples indicate that the LLM just searched the web and reported what it found? When you create a website with fake information and search Google for that information, it will of course bring up your site, not because it’s factually correct but because it’s related to what you searched for. What am I missing?
The part where lots of people have historically trusted LLM responses without verification, more than trying to sort through the dross on Google or Bing search results is, I think, the point.
The problem with this specific instance is that if you asked someone to find out who won this championship without using an LLM, they’d reach the same answer. I’d be much more impressed if someone managed to poison an LLM into answering that US won the 2023 World Cup
Ten years ago multiple tech giants openly stated they would not help the Trump administration build a Muslim registry [1]. Since then, several of them have bowed the knee and donated to his second inauguration.
I’m from Germany, and keep wondering how much more damage the NSDAP could have done if they had access to the data these companies now have on everybody.
[1] https://www.theverge.com/2016/12/16/13990234/google-muslim-r...
Again a reminder that Google etc have data far more than historical dictatorship like Stasi, KGB etc. just for better marketing. we need better data protection though GDPR is a good start.
Also: malware often tries to detect a VM or an emulator too, for example Windows Defender uses an emulator internally to detonate samples, and there are attempts by malware to detect this and change the behavior to something benign.
Way back in the early 90s Thunderbyte Antivirus' TBCLEAN would use the x86 trap flag to single step viruses up to the point where they restored the original entrypoint of the infected program, then write the "cleaned" program back to disk. They used the CPU single-step as a hack to alleviate needing to write an emulator.
The virus writer Priest figured out he could detect being run under single-stepping, and manipulate the stack and trap flag to re-vector control from TBSCAN to a destructive routine that trash the user's hard disk (but otherwise just run normally when not in the presence of TBCLEAN).
He later used this idea as the basis for the "emulating tracer" (in Natas, for sure-- but I think present in some earlier code too-- I don't remember what, thought) using single-step interrupt calls to trace thru resident antivirus programs to find original BIOS and DOS interrupt vectors and "call past" them (to prevent detection and do stealth).
His tracer decoded the next instruction to detect every method by which the trap flag state could be leaked to or mutated bu the traced code. He would emulate and step over any of these "privileged" instructions", presenting a sanitized state to the code under trace. It wasn't a full x86 emulator and could not have handled code that used trap-based anti-debug. That would have required a full emulator (and that way lies madness).
When VMware virtualized x86 I thought about Priest's code. Defender and other AV running samples under emulation makes me think about it too. So does this article.
Modern ransomware are not just encrypting data but uploading them somewhere too, the victim is then threatened with a leak of the data. A backup does not save you from that.
Well yes, if you get breached, you have problems. At least in good backups scenario you can continue to operate, so you have money incoming to fix this.
reply