I really wish these models were available via AWS or Azure. I understand strategically that this might not make sense for Google, but at a non-software-focused F500 company it would sure make it a lot easier to use Gemini.
1,713 karma · joined March 13, 2018
Also increasingly it's becoming important to look at token usage rather than just token cost. They say Opus 4.5 (with high reasoning) used 50% fewer tokens than Sonnet 4.5. So you get a higher score on SWE-bench verified, you pay more per token, but you use fewer tokens and overall pay less!
[0] https://www.anthropic.com/engineering/claude-code-best-pract...
How do you know this? Wouldn't we expect the benefits of AI in the legal industry to be way less likely to make the front page of HN?
Depending on volume, your total cost would likely be lower. I know you mentioned privacy concerns so this may not be an option. But it significantly simplifies your work, letting you focus on the parts themselves.
It's more complicated than that. Of the 6352 people who applied for TANF, 2306 dropped out during the process. Then of the 4046 TANF applicants remaining, only 2.6% tested positive for drugs. The vast majority of media coverage focused on the 2.6% being less than the ~8% drug-use rate in the general population.
What we don't know is of the people who dropped out, was this due to unintended reasons (privacy concerns, the inconvenience of the drug test, missing deadlines) or due to the intended reason (people self-selecting out because they knew they would test positive and become ineligible for 12 months). We'll never know the real breakdown, but it's misleading to say "it cost more than they saved".
As soon as I read this line I grimaced. This is a clear sign of an organization that doesn't respect peoples' time. The class should be an email (and proper follow-up by the management chain) establishing three rules:
- Meetings must have an agenda
- After a meeting, there must be a follow-up email describing what was decided and any action items
- Recurring meetings should be rare/exceptional
- Given good meeting notes and action items sent afterwards, reduce the invite list to decision makers; people who need to be informed can be added to the follow-up email
Big if true. Setting up an RL loop for training on math problems seems significantly easier than many other reasoning domains. Much easier to verify correctness of a proof than to verify correctness (what would this even mean?) for a short story.
I'm excited that this capability is getting close, but I think the current level of performance mostly makes for a good demo and isn't quite something I'm ready to adopt into daily life. Also, OpenAI faces a huge uphill battle with all the integrations required to make stuff like this useful. Apple and Microsoft are in much better spots to make a truly useful agent, if they can figure out the tech.
Instead, I think this is going to open a new paradigm with an immense long-tail of hyper-niche fit-for-purpose business applications. There's so much small-scale work that happens in corporations that isn't common enough to be worth building a product to solve. But it's still a big time-saving to the departments/users if they can improve the process with a vibe-coded app!
Doesn't seem like a very efficient way to accomplish the goal, but it would make for a great plot line in a spy movie.
It's possible, despite what many of these comments say, to build friendships as an adult. It just takes investment. What has helped me the most is:
- Bringing connections from one context into another. For example if you have a colleague you like, your relationship can't grow closer unless you start hanging out in a non-work context. Get involved with athletic activities with them. Or if you have similar aged kids, invite them over for a play date while the adults hang out.
- Combining friend hangouts. If you want 15 friends but only have time to hang with people every couple days, you don't have time to meet with people one-on-one. Organize group events like dinner parties or pot lucks. Also don't restrict these events to be just "work friends" or "pickle ball friends". Cross-pollinate groups.
- Create local friends. Nothing is more convenient than friends you live next door to. We do porch wine with our neighbors on Monday nights. It takes very little effort to set up a group text and get momentum around something like this.
OpenAI made a big show out of hiding their reasoning traces and using them for alignment purposes [0]. Anthropic has demonstrated (via their mech interp research) that this isn't a reliable approach for alignment.