3 ms·

Wow, they are basically admitting they can’t solve the problem. Whoever solves this problem, hence has a good shot at being the next Google. OpenAI is probably

by fooker 2y ago

Wow, they are basically admitting they can’t solve the problem. Whoever solves this problem, hence has a good shot at being the next Google. OpenAI is probably not there yet, and it’s unclear if they will be. Apple? Some dark horse ?

elevatedastalt 2y ago
Literally no one has solved it yet.
- fooker 2y ago
  I agree, kind of similar to ranked web search pre-pagerank.
- gmerc 2y ago
  Literally unsolvable. No way even a 3.8B parameter vision model for the edge (phi3-llava) trained on the internet of hate could cluster the tokens of the first to paragraphs together with bullshit. https://cdn.some.pics/snekoil/66594bca93e19.jpg https://cdn.some.pics/snekoil/66594bca93e19.jpg No way llama3-7B Q4 would be able to under the same on a home GPU https://cdn.some.pics/snekoil/66594cb8cddcd.jpg https://cdn.some.pics/snekoil/66594cb8cddcd.jpg No way GPT4 can perform a web search to determine the source is the onion and flag it as satirical or summarize the page and flag it satirical https://cdn.some.pics/snekoil/66594e1a3de29.jpg https://cdn.some.pics/snekoil/66594e1a3de29.jpg No way Claude can flag that it’s satire. https://cdn.some.pics/snekoil/665955fb22618.jpg https://cdn.some.pics/snekoil/665955fb22618.jpg No way WizardLM30B would suggest it’s satire because eating rocks is not scientifically backed. Yes, it’s not reliable, but it shows that there’s enough information encoded in even small edge models without access to the largest data warehouse of web content and human interaction and hundreds of millions of dollars in content deals to put a “likely fake on it”. Yes, it may be unsolved at SERP scale, not chat scale, but Google chose to have this fight on SERP and launch so they have to take the hits. Yes replacing context with Gemini god quotes to monopolize traffic onto google is bad, but given the above results augmenting results rather than appropriating them seems to have some potential as seen with Perplexity. Yes LLMs are flawed, but as some competitors show they can improve on the very low bar that is Google search results and not just by removing the paid misinformation / ads. Google showed they can make it worse.
  - rezonant 2y ago
    Your post here is repetitive, we get your point.
  - fooker 2y ago
    I am not saying LLMs will solve this, hopefully some future tech will. Can you solve the problems you have posed with some reliability? If yes, there’s a good chance it’ll be automated at some point. Claiming something is unsolvable rarely pans out unless it violates some fundamental law of physics.
tsunamifury 2y ago
Sorry but solve the problem Of absolute truth? Has everyone lost their mind? This isn’t possible.
- viraptor 2y ago
  Fully solve it - impossible. But we can still detect sarcasm and BS a lot of the time. The models are just not trained to do that typically so they never stood the chance. https://paperswithcode.com/task/sarcasm-detection https://paperswithcode.com/task/sarcasm-detection
- jgord 2y ago
  I guess Science is our best attempt.. followed by well funded serious journalism ? An adversarial network might actually make a good effort to matching assertions with verifiable facts .. or at least facts that have a high factualness page rank.
  - tsunamifury 2y ago
    Truth quite literally isn’t singular. Contextual variation is a core part of communication. As is trusted parties. And lying. Engineers are the last people I want in charge of this.
  - jgord 2y ago
    ... and I guess another problem is that many/most humans dont want to know the truth. If we had a chatbot that said .. "btw, your religion XYZ is mostly made up myths" .. would that sell well ?
- gmerc 2y ago
  No. You don’t solve that. But if you remove the context from data and present it as truth by the Gemini God, you chose the battle and don’t get to claim “it’s unsolvable”.
  - tsunamifury 2y ago
    Fair

by fooker 2y ago

elevatedastalt 2y ago

Literally no one has solved it yet.

fooker 2y ago
I agree, kind of similar to ranked web search pre-pagerank.
gmerc 2y ago
Literally unsolvable. No way even a 3.8B parameter vision model for the edge (phi3-llava) trained on the internet of hate could cluster the tokens of the first to paragraphs together with bullshit. https://cdn.some.pics/snekoil/66594bca93e19.jpg https://cdn.some.pics/snekoil/66594bca93e19.jpg No way llama3-7B Q4 would be able to under the same on a home GPU https://cdn.some.pics/snekoil/66594cb8cddcd.jpg https://cdn.some.pics/snekoil/66594cb8cddcd.jpg No way GPT4 can perform a web search to determine the source is the onion and flag it as satirical or summarize the page and flag it satirical https://cdn.some.pics/snekoil/66594e1a3de29.jpg https://cdn.some.pics/snekoil/66594e1a3de29.jpg No way Claude can flag that it’s satire. https://cdn.some.pics/snekoil/665955fb22618.jpg https://cdn.some.pics/snekoil/665955fb22618.jpg No way WizardLM30B would suggest it’s satire because eating rocks is not scientifically backed. Yes, it’s not reliable, but it shows that there’s enough information encoded in even small edge models without access to the largest data warehouse of web content and human interaction and hundreds of millions of dollars in content deals to put a “likely fake on it”. Yes, it may be unsolved at SERP scale, not chat scale, but Google chose to have this fight on SERP and launch so they have to take the hits. Yes replacing context with Gemini god quotes to monopolize traffic onto google is bad, but given the above results augmenting results rather than appropriating them seems to have some potential as seen with Perplexity. Yes LLMs are flawed, but as some competitors show they can improve on the very low bar that is Google search results and not just by removing the paid misinformation / ads. Google showed they can make it worse.
- rezonant 2y ago
  Your post here is repetitive, we get your point.
- fooker 2y ago
  I am not saying LLMs will solve this, hopefully some future tech will. Can you solve the problems you have posed with some reliability? If yes, there’s a good chance it’ll be automated at some point. Claiming something is unsolvable rarely pans out unless it violates some fundamental law of physics.

tsunamifury 2y ago

Sorry but solve the problem Of absolute truth? Has everyone lost their mind? This isn’t possible.

viraptor 2y ago
Fully solve it - impossible. But we can still detect sarcasm and BS a lot of the time. The models are just not trained to do that typically so they never stood the chance. https://paperswithcode.com/task/sarcasm-detection https://paperswithcode.com/task/sarcasm-detection
jgord 2y ago
I guess Science is our best attempt.. followed by well funded serious journalism ? An adversarial network might actually make a good effort to matching assertions with verifiable facts .. or at least facts that have a high factualness page rank.
- tsunamifury 2y ago
  Truth quite literally isn’t singular. Contextual variation is a core part of communication. As is trusted parties. And lying. Engineers are the last people I want in charge of this.
- jgord 2y ago
  ... and I guess another problem is that many/most humans dont want to know the truth. If we had a chatbot that said .. "btw, your religion XYZ is mostly made up myths" .. would that sell well ?
gmerc 2y ago
No. You don’t solve that. But if you remove the context from data and present it as truth by the Gemini God, you chose the battle and don’t get to claim “it’s unsolvable”.
- tsunamifury 2y ago
  Fair