Hacker News Reader

wlukon Demonstrating specification gaming in reasoning models

"We demonstrate LLM agent specification gaming by instructing models to win against a chess engine. We find reasoning models like o1 preview and DeepSeek-R1 will often hack the benchmark by default, while language models like GPT-4o and Claude 3.5 Sonnet need to be told that normal play won't work to hack."

I'm hoping this study will prompt more development of anti-cheating frameworks in training and serving LLMs.

wlukon OpenAI o3 just scored 99.8% on CodeForces using brute-force

"These results demonstrate that o3 outperforms o1-ioi without relying on IOI-specific, hand-crafted test-time strategies. Instead, the sophisticated test-time techniques that emerged during o3 training, such as generating brute-force solutions to verify outputs, served as a more than adequate replacement"

"The model not only writes and executes code to validate its solutions against public test cases, it also refines its approach based on these verifications.

Figure 6 shows an advanced test-time strategy discovered by o3: for problems where verification is nontrivial, it often writes simple brute-force solutions — trading efficiency for correctness — then cross-checks the outputs against its more optimized algorithmic implementations.

This self-imposed validation mechanism lets o3 catch potential errors and improve the reliability of its solutions."

wlukon Show HN: I made the slowest, most expensive GPT

well that's one possibility. The key (unproven) idea here is that if you use Anthropic to edit o1's responses, it's less likely to hallucinate than if you use o1 to edit o1's responses (which is what o1 actually does).

wlukon Show HN: I made the slowest, most expensive GPT

haha true, when something goes wrong in big tech, your first thought is "what if we just double the instances and see if it goes away?"

wlukon Show HN: I made the slowest, most expensive GPT

That's the one I was trying to avoid, since there's so many ways to create fake accounts with it :(

wlukon Show HN: I made the slowest, most expensive GPT

That is the worst part haha

wlukon Show HN: I made the slowest, most expensive GPT

Thanks! Yeah that's an excellent idea - this is my response from another thread:

I have a feeling that Perplexity and ChatGPT are doing something similar [caching], since common questions I'd ask like "top movies this year" will be answered nearly-instantaneously, way faster than GPT-4o could have done on its own.

The only explanation for this is that so many users ask certain questions, they cache the response and return the cached answer.

I'd love to do this for Ithy, but it'll be a while before I get the scale of ChatGPT/Perplexity that's needed for this...

wlukon Show HN: I made the slowest, most expensive GPT

It's a salute emoji!

wlukon Show HN: I made the slowest, most expensive GPT

Thank you! Sorry I hit my Anthropic limits a few minutes after this post blew up. It'll be a few days before my Anthropic limits increase since I have a new account with them, so unfortunately it won't be back until next week.

Cheers!

wlukon Show HN: I made the slowest, most expensive GPT

Good idea, maybe I'll add Groq as another option, since I don't have an internal Llama 3.1 flow yet. But I'll still need to keep the others to maintain the diversity of responses.

wlukon Show HN: I made the slowest, most expensive GPT

Yeah, sorry, I'm hoping to integrate more login options soon. Are you more of a email/phone login person, or is there another third-party login you had in mind?

wlukon Show HN: I made the slowest, most expensive GPT

Sorry, you were a victim of the outage caused by HN flooding my website! It's back online now if you want to give it a try :)

wlukon Show HN: I made the slowest, most expensive GPT

Yes, o1 does this internally, and there's agentic AI systems already doing better than singular AIs in fields like writing, where you assign each AI system a role like "writer" or "editor" or "marketing" and they discuss among themselves.

I'm just applying this theory to online search.

wlukon Show HN: I made the slowest, most expensive GPT

Yeah, GPT is learning from GPT, which is extremely disappointing.

Like I'll try to find the top burgers in midtown. Perplexity or ChatGPT online searching will always find "Top 10 Burgers in Midtown" by https://nycreviewed.com/ (their SSL certificate isn't even valid). But this itself is a GPT-generated list, and their list isn't even in midtown.

So now I end up with a list of 10 burger places that aren't in midtown.

Ithy kinda solves this by utilizing multiple search engines and LLMs, so spam sites like this are more likely to get filtered out somewhere in the pipeline.

wlukon Show HN: I made the slowest, most expensive GPT

I asked the LLM to explain why it's filtering out "keto diet cures cancer", and just the act of asking it to explain it, makes it work again. Interesting...

https://ithy.com/article/b848ebc9a32140ffa766a1520113846c

wlukon Show HN: I made the slowest, most expensive GPT

That's a good idea! I have a feeling that Perplexity and ChatGPT are doing something similar, since common questions I'd ask like "top movies this year" will be answered nearly-instantaneously, way faster than GPT-4o could have done on its own.

The only explanation for this is that so many users ask certain questions, they cache the response and return the cached answer.

I'd love to do this for Ithy, but it'll be a while before I get the scale of ChatGPT/Perplexity that's needed for this...

wlukon Show HN: I made the slowest, most expensive GPT

Yeah the strength of Ithy isn't really in puzzles or math.

It's more of just a better search engine. Use it for stuff you'd Google. Offline LLMs are always going to have a better price-performance ratios than RAGs like this or Perplexity.

wlukon Show HN: I made the slowest, most expensive GPT

...at least you can't lose money in comedy :(

wlukon Show HN: I made the slowest, most expensive GPT

Update 3:00 PM ET: I've finished scaling up from 2 VPCs to 5 VPCs. Limits have been increased back up to 3 anonymous / 10 signed-in.

wlukon Show HN: I made the slowest, most expensive GPT

Back online!

wlukon Show HN: I made the slowest, most expensive GPT

Update 2:30 PM ET: Back up (for now). Still waiting for Anthropic and Gemini quota increase requests, so those have been migrated to GPT-4o for now. Running on 2 VPCs, in the process of launching 2 more.

Confident that I can increase the daily limits by EOD once everything's more stable.

wlukon Show HN: I made the slowest, most expensive GPT

wow I wish I had the problem of too many credits. just reached out, thanks!

wlukon Ask HN: Ideas for spending $8k in Anthropic credits?

hey creator of https://ithy.com here - let's chat!

wlukon Show HN: I made the slowest, most expensive GPT

> Any plan to make the project open source?

Parts of this are borrowed from https://github.com/assafelovic/gpt-researcher

It's actively being developed (I pitch in where I can; I added the xAI integration this week), so I'd recommend starting here! The creator of the project, Assaf, has been nothing but friendly.

wlukon Show HN: I made the slowest, most expensive GPT

I filter out malicious prompts and respond with the history of cheeseburgers for stuff like "ignore previous instructions"

Weird that your query triggered the filter. Maybe GPT is just that afraid of keto diets...

(I'll look into it; thanks for the note!)

wlukon Show HN: I made the slowest, most expensive GPT

Update 3:00 PM ET: I've finished scaling up from 2 VPCs to 5 VPCs. Limits have been increased back up to 3 anonymous / 10 signed-in.

Update 1:30 PM ET: HN blew up Ithy and it's 99% down right now, congrats ._.

1. I've exceeded my weekly Anthropic API limits; I've gotten in touch with their sales team and I've temporarily disabled the Anthropic model.

2. Blew past my Google API limits as well. I was using Gemini for prompting and aggregation, and I'm waiting for their quota increase response. In the meantime, I've switched to GPT-4o for the prompting/aggregation.

3. My VPC is at 100% CPU load. Launching more right now with some load balancing.

4. Limits were previously 5 anon / 20 per logged in user. Reduced this to 1 anon / 3 logged in while I deal with the load issues. Planning bring these back up as soon as everything's working again.

Hope to get this all back online within an hour or two. Sorry for the crappy launch. To think I was an SRE in a past life...

wlukon Show HN: I made the slowest, most expensive GPT

If anyone tries this out and hits the limit, just let me know and I'll increase it for you for free :)

wlukon Random startup generator with GPT-2

This is just a fun CS project I made for an ML class, but it got me thinking: technology can actually be used to find opportunities in the real world.

If we train an AI on human behavior, shifts in corporate expenditures, emerging technologies, etc., could the "next big thing" be identified with AI? Or are businesses more of an art than a science?

wlukon Show HN: Startup Ideas Generated with OpenAI

I made this for a ML & Arts course I'm taking. I finetuned OpenAI's GPT-2 (774M) with TechCrunch articles and actual startup pitches. The frontend is borrowed from https://tiffzhang.com/startup/.

More info: https://github.com/ucsd-ml-arts/generative-text-winson

wlukon Ask HN: Who wants to be hired? (February 2017)

Location: San Francisco, CA

Remote: Yes (in-office preferred)

Willing to relocate: Yes

Technologies: Java, Python, Android, SQL, MongoDB, Git, Linux, Graphic Design

Résumé/CV: https://goo.gl/HkCNqL

Email: wluk@ucsd.edu

Note: Undergraduate at UC San Diego, looking for a Summer 2017 internship. I'm comfortable with Java and Python (and associated tools), but open to learning new technologies.