Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Anthropic Launches Claude Opus 4.8 With Gains in Coding and Honesty
- Thread starter MacRumors
- Start date
- Sort by reaction score
Crowbot
macrumors 68000
I usa Claude a lot but only for informatioanl questions. I don't want to get chummy with it. I like it because, if i ask the questions carefully enough, I can get deeper answers to the question. And I cought a few mistakes. Like it didn't know there is a mag stripe on the back of the Apple Card. For me, it's mostly a souped-up DDG.
southerndoc
Contributor
I wish there was a way to default having web search enabled by default. There is either a bug or a design flaw with the macOS desktop app where Claude constantly tells me it's limited in its ability to search the web.
Spec360
Suspended
HouseofG
macrumors 6502a
BSDnostalgia
macrumors 6502a
And yet, legally, it is a foreseeable consequence that some people will forget that. We arrive, again, at a legal impasse when people mistake AI bots for persons.
MacATDBB
macrumors regular
4.8 comes across to me as slightly odd. 4.7 was solid as long as you pushed back when it got silly. But 4.8 seems almost neurotic about its own honesty whilst simultaneously failing to flag terms it doesn't understand and instead assuming they mean something else. I may skip this version as 4.7 remains a non-default optional model.
[Update] After working more with 4.8, no more oddness has occurred and the coding seems to have improved.
[Update] After working more with 4.8, no more oddness has occurred and the coding seems to have improved.
Last edited:
ChamomileTea
macrumors member
OK then.
When office drone begin losing their jobs despite companies maintaining their revenue, we'll see that…
usmaak
macrumors 65816
Honesty. 😂 It just means that it is less likely to hallucinate. I doesn't actually lie. It truly believes everything it tells you until you point out its errors. People treat this like it is infallible. If you're truly working well with it, you are always assuming that it is wrong until you can prove it right. It's not human. It's a very clever prediction engine but that's really all it is. It's a tool.
verniesgarden
macrumors 65816
honestly think it's going to even itself out. there was tons of doom and gloom about the internet when it started.... wait never mind I take back what I said... we are screwed.
verniesgarden
macrumors 65816
Jumpthesnark
macrumors 68000
Good to see this. AI is improving at a fast pace. Will try out the new model!
I'm doing my best to keep pace with the advances in AI model evolution, but man oh mighty! It's hard to keep up!
I'm curious how they work to improve the characteristics? Technological improvements enhance speed, but what are they doing differently to improve honesty? Catching more edge cases?
I'm curious how they work to improve the characteristics? Technological improvements enhance speed, but what are they doing differently to improve honesty? Catching more edge cases?
No, because the world is not black and white. This type of thinking reminds me of what I see in the ******** landscape where one stands outside, notices the temperature being nice and cool, and says there's no such thing as global warming. It's not a "this or that" thing, but scales. Honesty also exists in varying degrees. Someone can be "mostly honest" about something, slipping in a little white lie here and there. We've all done it.
Furthermore, just because a response may not be fully honest doesn't automatically mean that the model is being nefarious. Again, not a this-or-that situation.
How will it enhance our lives? How many people are getting into technology because of it? How many new jobs are being created to build infrastructure, build software and maintain what is built?
The world evolves. Those that hang on to the past are doomed to suffer the consequences.
If there was no such thing as evolution, we would not exist.
iPay
macrumors 6502
Jumpthesnark
macrumors 68000
Someone is "honest" or not. Sorry, but I don't buy the sliding scale idea of honesty. There is a word for being "most of the way" to honest, but not being honest. That word is "dishonest."
If this company or the AI world in general thinks there's a sliding scale for honesty, then I think they're using the wrong word.
How about "trustworthy?" "Credible?" Both of those accurately reflect the user's ability to judge the work the AI has done, rather than "honest," which implies that a string of code knows the difference between the truth and a lie, or can care.
But if it's about how an answer makes you feel, and not whether it's accurate, there's a word for that too.
Yr Blues
macrumors 68040
Until a data center moves into your neighborhood. Your job is lost. Your kids become dumber. Your bills go up. Your data surveilled. Progress! Everyone that opposes AI are terrorists.
It astounds me how many people have critical hot takes without a) reading the freaking article which takes like 2 minutes, or b) having any recent experience with the models. Humans can sure produce a lot of slop too.
Claude code would occasionally lazily implement although 4.7 was better at direction following vs. 4.6 which was a bit better at general use (much better for chat for ex.). 4.8 seems good at chat again, and the effort toggle is welcome.
I still lament the loss of voice mode with the best models. So, so stupid they cut that.
I'm interested in the needle bench results but haven't seen those – I actually did not run into much problem with long contexts but I do kind of refresh code's memory when I run long. Many of my sessions run for 1M context regularly, I just don't do critical work with the last ~150k and use it to handoff.
Claude code would occasionally lazily implement although 4.7 was better at direction following vs. 4.6 which was a bit better at general use (much better for chat for ex.). 4.8 seems good at chat again, and the effort toggle is welcome.
I still lament the loss of voice mode with the best models. So, so stupid they cut that.
I'm interested in the needle bench results but haven't seen those – I actually did not run into much problem with long contexts but I do kind of refresh code's memory when I run long. Many of my sessions run for 1M context regularly, I just don't do critical work with the last ~150k and use it to handoff.
