00:00 Speaker A
Well, I have a question about this new AI model from Anthropic Dan. Please fill in. What do you need to know?
00:04 Dan
Yes, this is Sonnet 5. This is likely the third model of the 5 Series. They obviously have Mythos 5. This is their huge model and then Fable 5. It’s like a model based on Mythos but with more security features. Both of these are still bound by government regulations. The government has given some companies access to them, but that’s another story. This is basically considered to be a cheaper version of their model, rather than something like the previous generation’s top-of-the-line model, the Opus 4.8.
01:03 Dan
What they’re actually trying to do here seems to be leaning towards lowering the cost per token when it comes to AI. Tokens are essentially a unit of measurement for AI. Tokens can be anything from short phrases to word fragments. The idea here is that if costs can be reduced, companies will be less likely to try to stop using AI as often. We’ve also seen some companies reduce spending on AI. Meta mentioned it, but Uber has breached its own cap on AI spending. Amazon Woo will remove the leaderboard of how many Woo tokens people are using. It seems like the whole era of this kind of token maximization is in the rearview mirror. So companies like Genpic are responding with more efficient and lower cost-to-use models.
02:02 Speaker A
Dan, which jobs could this new AI model impact? In your opinion, which professions should be paying close attention?
02:12 Dan
In general, I think it’s overall, well, kind of a normal model. I don’t think anything can really be normal when it comes to AI, but I would say it’s a threat, not a specific job, and overall. This is a more agent-like use case. So, um, it’s like a place to actually sit. Well, one is not better than the other when it comes to their overall usage in different fields. This is a more common model. Well, it’s good for the agent, and like I said, it’s cheaper for that use case.
03:00 Speaker A
Dan, are hallucinations still a problem? In other words, can these AI agents be trusted to work independently?
03:08 Dan
Yeah, I mean, there’s always the fear of hallucinations, right? So no matter how good these models are, they remain in the background. You know, they’ve made, uh, improvements to alleviate those kinds of issues, but the error still occurs. So when it comes to doing a simple survey at a point in time, uh, um, we were doing a survey to find out the dates of various earnings reports for the last quarter, and uh, Gemini was providing completely wrong dates. Oh, and, you know, here and there things happen along those lines. But I think it has gotten better. When it comes to AI, I really do. If you are, we are obviously journalists, so of course we have to double check our work every time. But I think everyone who uses them should care.
