Anthropic Launches Claude Opus 4.8 With Gains in Coding and Honesty

[Agent007]

So this is Mac Rumors isn’t it? Why does this site have so much coverage of minor AI model updates?

they're paid by Big Tech to push slop, of course!

 

[jimbobb24]

Anthropic is the hypiest hype machine. Every few months some hype nonsense. But they do good work for coding but I prefer Codex. But my son and I just wrote an app for home use with Claude.

 

[jimbobb24]

I'm more and more realizing that the reason many people are treating AI like they would a person (going so far as to ascribing consciousness to them) is because they basically see other people as intelligent machines. Sad state of affairs.

If you listen to some of the AI gurus they will say AI is conscious. And if you dig into their writings often you will discover exactly what you said...they think human consciousness is a bunch of algorithms without insight just following a chemical path. So sure AI is as conscious as the normal human but neither would be conscious under most peoples definitions.

 

[everlast3434]

I'm holding out for 4.9!!!

 

[dropadrop]

Quantity over quality. That’s pretty much the definition of slop.

Generated images all have the same feel, code has ridiculous bugs, writing comes out bland. I’ve tried. After fighting with Claude to get the code right, I gave up and found it faster to just write the code myself. The only people who don’t consider it slop are those who don’t know any better.

The fact that you can do stupid things with a tool does not mean you can’t also do really amazing things with it as well.

The current hype is pushing a lot of people to ”do something” with AI without a good understanding of what it’s good at, or how it could really help. That leads to a lot of slop being released, but that does not mean the technology itself is bad.

I have 25 teams of developers working on my products. We’ve been using different AI models increasingly for 15 years and now part of the problems devs are managing to solve with LLM would have been close to impossible (with our resources) to solve before.

However even I got ”do some features with AI” requests from higher leadership a few years ago. To end up with stuff that is not slop you start with the problem you want to solve and then pick the tools and technology to solve it - not the other way around. But blaming the tools because there are so many bad carpenters around just means you are not a carpenter (at least in this case because Claude is not slop).

 

[eifelbube]

They do have actual intelligence. What they don’t have is continual (online) learning. That will be solved at some point but they have to find an energy efficient way about that.

Emotional intelligence too? Intelligence is not one-dimensional — luckily.
Aside from many other benefits (e.g. brain training), continual offline learning by humans keeps the electricity bill and water consumption lower.

 

[Happy_John]

You are making some fair points, but I think your analogy breaks down a bit. Note that I'm not saying you're incorrect, I'm simply suggesting some things to think about.

I agree that current LLMs do not comprehend in the human sense. They do not have subjective awareness, lived experience, etc. They can produce language about pain, grief, or morality without experiencing any of those things as humans do. To be fair, people can do the same thing -- rationally talk about things they have not personally experienced, but such comments come from a place without phenomenological understanding. However, saying an LLM is “just finishing a story...based on statistical probability” already defines its output as non-comprehending before establishing what comprehension is. You are building your conclusion into your premise.

True, but it’s a forum, and for a technical forum, not that technical or formal. This is entertainment, not writing an abstract for a journal. Honestly, it would be a ware of time putting the effort into developing a formal argument on here. There are more appropriate places for that.

Also, distilling this down to an equivalence to someone singing English lyrics phonetically without understanding the words is going a little far in the reduction. The singer cannot flexibly explain the song, translate it, critique it, apply its themes to another context, or revise it for a different audience. LLMs can do many of those things. That does not prove consciousness or true understanding, but it does suggest something more complex than rote mimicry.

That’s closer to where I’m going, except saying “something more complex than rote mimicry” implies I think it is “just rote mimicry”. It is. Don’t underestimate the power of taking in large amounts of data, then using that data to develop both a reasoned output, but more importantly, developing a method of reasoning itself.

Reasoning does don’t develop in a vacuum and then is applied to data. Reasoning is developed through a huge amount of data input. The input precedes the development of reason8ng.

There’s the old trope of people talking about their education as a huge amount of information that was not in any way relevant to their adult life. That’s not completely true, but a lot of that educational input is to enable the development if critical thinking and reasoning.

Yes LLM based models can reason. I’m not underplaying that. But that reasoning is a result of inputting and training based on statically probabilities. Reason is the result of analysis Ing data, it isn’t separate to it.

Reason, human or otherwise, itself could be described as a „story completer”: examine data, deduce patterns, extrapolate. It is possible to use formal logical reasoning to reach a conclusion, while not necessarily comprehending the content of the problem.

Reason is not comprehension. When we think of „human intelligence”, it is not simply reasoning.

This will continue to be philosophical. The better distinction about what's happening may be between phenomenological understanding (going back to my second paragraph) and functional language-based modeling. Humans integrate language into lived experience, emotion, goals, social behavior, and self-awareness. LLMs do not (there's an assumption there, but it's a safe one at this stage of development). However, the LLMs can still model contextual semantics well enough to generate useful responses.

So I agree that we should not mistake fluent output for human comprehension, wisdom, or authority. But I would not reduce LLMs to mere karaoke. They are not minds as we typically think of minds, but they are also not simply just repeating things. They are sophisticated systems that can simulate aspects of reasoning through language, while lacking the consciousness that human understanding depends on (again, another assumption in there!).

Again, I’m NOT saying “mere karaoke”. In my anecdote, I wasn’t belittling those singers. Was it fake. No. Did they give excellent renditions of the songs. Yes so did they achieve their aim. Yes. But they achieved that aim through a process very different to the process people would except. They achieved that process without comprehension ( u derstanding of the greater context behind a query / instruction ).

The kanged used in the original reticle “honesty” is unhelpful. An AI simply gives the prompter what he/she asked for. If this means fabricating / hallucinating non-facts that appear convincing, then it has achieved its aim. In a sense, you can’t criticise it. It is not doing this to be edgy or lazy, it is doing what it was asked.

It’s not a person, it’s a nomreflexive genie. It fulfills requests, and it will fulfil that request with fabrication, if it needs to.

It’s not that the karaoke singers are lesser, it’s that a listener would assume they knew the some through comprehension, but this was not th case.

In the same way, people as making misassumptions about „intelligence” of AIs, because the output may have the appearance of comprehension (I.e. consciousness and self-awareness), when they AI has been able to price this output without any self-awareness or comprehension.

To try and sum it up, humans inevitably ask “cui bono?”, AI does not.

If is is “intelligence”, it is not analogous to “human intelligence”, even if it reaches the same conclusion as “human intelligence” in answer to a specific query.

 

[tekboi]

Great! Now I can blow through my usage even faster!

 

[eifelbube]

Testers have found Opus 4.8 to be "more reliable and sharper in its judgement" when doing agentic tasks, and the model also made gains in honesty.​

Good to know this. Waiting for more improvements in honesty … while continuing not to subscribe to consumer AI because it’s been mganu:
- manipulative -
- gullible -
- absurd -
- noxious -
- unethical -

 

[macshmack]

what is this twaddle,

Rates of misaligned behavior like deception are lower than Opus 4.7​

I remember they marketed Opus 4.7 as 'low on acting in your best interest, high on deception (honestly!)'

 

[wilhoitm]

So this is Mac Rumors isn’t it? Why does this site have so much coverage of minor AI model updates?

Because Siri 2 is coming and it is going to be everything you hoped and dreamed it would be! 🤣

 

[spazzcat]

Good luck having enough tokens to do anything with it on even the Pro plan.

In just three days, I successfully developed a new add-on for our nonprofit website. This add-on integrates seamlessly with the backend it built plus thrid party APIs, including credit card processing. This included serveral hours of debugging had no issues with tokens.

 

[MacGod]

Claude is great but OPUS comes at a cost. It should only be used for research intensive operations and complex calculations. We advise users that most of what they want to accomplish with Claude can be done with Sonnet and to only use OPUS when the results of their output isn't as expected.

 

[tjminter]

In just three days, I successfully developed a new add-on for our nonprofit website. This add-on integrates seamlessly with the backend it built plus thrid party APIs, including credit card processing. This included serveral hours of debugging had no issues with tokens.

could be a skill issue for me. I use it for legal work.

 

[diblasio]

I hope it's more honest about it eats up it's tokens. Claude can burn through tokens with a simple task because of how it iterates and I find I'm always bumping into limits unless I am really specific with my prompt.

Be more specific. Not a big deal.

 

[diblasio]

When one of the most popular AI companies has to brag about improving their AI tool's honesty, you know we are all doomed. People already believe everything AI tells them, and probably 70% or more of it is pure garbage information on top of bias and lies. And without any government regulation—which there seems to be no plans for any to be implemented in the US anyway—it's the Wild West out there. I've seen so much AI slop online lately it's sickening. It's getting worse every day, and soon we won't even be able to differentiate AI slop from reality. Scary times, folks...

Because before AI people were well read and the government was honest! 🥴

 

[AeroEd]

No, because it is a tool, not a person. The responsibility lies solely with the person using it.

If you accidentally killed someone with your car or rifle, the car or rifle would not be responsible, you would be.

A rifle doesn’t lie, it means what it says?

 

[Happy_John]

A rifle doesn’t lie, it means what it says?

???? A badly made gun could explode in your hand and take most if your atm off. A gun below MG played with by a child might kill the child. But the gun itself is not displaying morally bad behaviour. That’s the extension if the analogy.


An LLM AI is a tool. Saying it lies” is anthropromising it. It is not a person.

An LLM AI’s purpose it to respond to your query and give you what you asked for. That it may do this by fabricating “facts” when it cannot access actual facts is a serious problem, but framing it as lying isn’t helpful - again, not a person, so stop ascribing human behaviours to a tool.

As pointed out many times, the responsibility lies solely with humans; it’s fir a court to decide in a specific case whether the responsibility vilify lies with the user fir taking the output at face value, or the company selling access to the model that provided the output.

Atop thinking that AI is anything other than an area if technology - it’s neither good or bad, but it can bloodily be used in makucoud ir careless ways, just as any technology can.

Blaming the AI is essentially an excuse to enable humans to avoid responsibility.

 

[CalMin]

Be more specific. Not a big deal.

My comment was a little flippant, but sure - here's what I run into constantly with Claude.

I use the $20/month Claude Pro. When I throw it a problem consisting of perhaps a couple of large spreadsheets, it can iteratively work on the problem then tell me I've reached usage limits for the session. Then it stops until whenever the limits reset.

I get around it by writing Python scripts to parse the spreadsheets first, then have Claude work with more compact JSON files. Still - it's not something that I ever ran into with ChatGPT.

 

[Wahlstrm]

They do have actual intelligence. What they don’t have is continual (online) learning. That will be solved at some point but they have to find an energy efficient way about that.

Ask any LLM about honesty and intelligence and they will literally all tell you they lack any form of meta cognition and therefore can’t recognise when they make stuff up or jump to weird conclusions.

They don’t even know the meaning of the words they use, it’s just a bunch of characters with a high statistical probability of being used together by humans.

 

[iMac The Knife]

Because before AI people were well read and the government was honest! 🥴

And I claimed neither of those. AI is only making it worse.

 

[BSDnostalgia]

Ask any LLM about honesty and intelligence and they will literally all tell you they lack any form of meta cognition and therefore can’t recognise when they make stuff up or jump to weird conclusions.

They don’t even know the meaning of the words they use, it’s just a bunch of characters with a high statistical probability of being used together by humans.

I just now asked GPT-5 if "you", that is, (GPT-5) is honest, and it answered:

"Yes — I aim to be honest: I provide accurate information, cite sources when I use web search, and clearly state limits or uncertainties."

"AI" isn't "a person", it is a massive interconnected mess of (technical term here🙂 "black boxes") that have been "trained" on the collected works of humanity and configured to answer questions. A fear that many of us have is that even simple distinctions regarding truth may get lost in the clouds of data, and, "AI" may answer a legal question citing "Jarndyce v Jarndyce". But, worse, when it advances to the point where it won't make that mistake, it still will have too much data and not enough reality.

 

[Happy_John]

They do have actual intelligence. What they don’t have is continual (online) learning. That will be solved at some point but they have to find an energy efficient way about that.

You're back to the same problem. Define "intelligence" ( actual or otherwise). AIs do not possess "human intelligence", so you can't say if AIs possess "intelligence" until you have a very clear idea of what "intelligence" is.

Impersonating being human or impersonating personality is not itself an indicator of intelligence.

AIs do NOT have personality. They have been trained to mimic having personality simply to appeal to users.