> But we know that any person who uses AI is likely to improve at what they do.
Do we?
I could have sworn there was research that stated the more you use these tools the quicker your skills degrade, which honestly feels accurate to me and why I've started reading more technical books again.
> I've started reading more technical books again
How's that working out for you in the context of working with AI tools? Do you feel like it's helping you make better use of them? Or keeping your mind sharp?
I've been considering getting some books on core topics I haven't (re)visited in a long time to see if not having to write as much code anymore instead gives me time to (re)learn more and accelerate.
I would suggest that any person who uses AI will atrophy their compositional skills unless they specifically take care to preserve those skills.
As a student, I constantly worry about this. But everyone in my class is producing output at a pace I can't compete with without AI assistance.
My brother is a CS student and he is pretty much in the same boat.
what class are you in that "producing output at a [rapid] pace" is relevant to the grade?
pick any cs class
I have a minor in CS and no -producing the assignment by the deadline is important- grades are not based on quantity of code vs classmates.
I mean, maybe things have changed (I finished college about 20 years ago), but I don't remember producing large volumes of stuff as being a particularly important part of a CS degree.
Between a challenging job market, increasing new frontiers of learning (AI, MLops, parallel hardware) and an average mind like mine, a tool that increases throughput is likely to be adopted by masses, whether you like it or not and quality is not a concern for most, passing and getting an A is (most of my professors actively encourage to use LLMs for reports/code generation/presentations)
It will be a very interesting experiment when your generation of computer science graduates enters the job market, to put it mildly.
That was never a worry in any of my CS classes.
Copying AI slop isn’t producing output! It’s also not conducive to learning
Yah and this seems to be supported by preliminary evidence on the impact of AI on things like retention and cognitive ability.
Not until large-N research is done without sponsorship, support, or veiled threats from AI companies.
At which point, if the evidence turns out to be negative, it will be considered invalid because no model less recent than November 2027 is worth using for anything. If the evidence turns out to be slightly positive, it will be hailed as the next educational paradigm shift and AI training will be part of unemployment settlements.
Clearly this means Anthropic believes this but would be nice to have a footnote pointing to research backing this claim.
We DEEPLY do not.
That's not, IMO, a "skills go down" position. It's respecting that this is a bigger maybe than anyone in living memory has encountered.
Let me add a single data point.
> is likely to improve at what they do
personally, my skills are not improving.
professionally, my output is increased
My software development skillset has improved. I’m learning and stress testing new patterns that would have taken far longer pre-AI. I’m also working in new domains and tech stacks that would have taken me much longer to get up to speed on.
People who use AI mindfully and actively can possibly improve.
The olden days of buidling skills and competencies are largely dying or dead when the skills and competencies are changing faster than skills and competency training ever intended to.
If things change fast, learning becomes even more important. And learning about the principles that don't change becomes most important of all.
I would even say it's likely the opposite. My output as a programmer is now much higher than before, but I am losing my programming skills with each use of claude code.
So I guess the key takeaway is basically that the better Claude gets at producing polished output, the less users bother questioning it. They found that artifact conversations have lower rates of fact-checking and reasoning challenges across the board. That's kind of an uncomfortable loop for a company selling increasingly capable models.
> the less users bother questioning it
This makes me think of checklists. We have decades of experience in uncountable areas showing that checklists reminding users to question the universe improve outcomes: Is the chemical mixture at the temperature indicated by the chart? Did you get confirmation from Air Traffic Control? Are you about to amputate the correct limb? Is this really the file you want to permanently erase?
Yet our human brains are usually primed to skip steps, take shortcuts, and see what we expect rather than what's really there. It's surprisingly hard to keep doing the work both consistently and to notice deviations.
> lower rates of fact-checking and reasoning challenges
Now here we are with LLMs, geared to produce a flood of output which strikes at our weak-point, the ability to do intentional review in a deep and sustained way. We've automated the stuff that wasn't as-hard and putting an even greater amount of pressure on the remaining bottleneck.
Rather than the old definition involving customer interaction and ads, I fear the new "attention economy" is going to be managing the scarce resource of human inspection and validation.
Sounds like having a strong checklist of steps to take for every pull request will be crucial for creating reliable and correct software when AIs write most of the code.
But the temptation to short change this step when it becomes the bottleneck for shipping code will become immense.
> So I guess the key takeaway is basically that the better Claude gets at producing polished output, the less users bother questioning it.
This is exactly what I worry about when I use AI tools to generate code. Even if I check it, and it seems to work, it's easy to think, "oh, I'm done." However, I'll (often) later find obvious logical errors that make all of the code suspect. I don't bother, most of the time though.
I'm starting to group code in my head by code I've thoroughly thought about, and "suspect" code that, while it seems to work, is inherently not trustworthy.
I think we're still at the stage where model performance largely depends on:
- how many data sources it has access to
- the quality of your prompts
So, if prompting quality decreases, so does model performance.
Seems like it’s impossible for output to be good if the prompt is bad. Unless the AI is ignoring the literal instructions and just guessing “what you really want” which would be bad in a different way.
Sure, but the study is saying something slightly different, it's not that people write bad prompts for artifacts, they actually write better ones (more specific, more examples, clearer goals,...). They just stop evaluating the result. So the input quality goes up but the quality control goes down.
What does prompting quality even mean, empirically? I feel like the LLM providers could/should provide prompt scoring as some kind of metric and provide hints to users on ways they can improve (possibly including ways the LLM is specifically trained to act for a given prompt).
That would be a quality metric, and right now they are focused on quantity metrics.
I’m not alone in finding this against the claims of the product right?
Claude is meant to be so clever it can replace all white collar work in the next n-years, but also “you’re not using it right?” Which one is it?
Anthropic is a weird company where the CEO almost admits at times they are probably building the Torment Nexus, yet still feel they need to do it anyway…because someone else might do it first?
Which one will convince you to buy more Claude? Please answer honestly, it's for the sake of profits.
Anthropic in particular seem to be in a weird place where on the one hand they fund some real research, which is often not all roses and sunshine for them, but on the other hand, like all AI companies, they feel the need to make absurdly over-the-top claims about what's coming up Real Soon Now(TM).
I'm not quite convinced of the maximalist claims, but these two aren't incompatible. Every time we talk about a company being "mismanaged" by e.g. a private equity buyout, what we mean is that the owners had access to a large volume of high quality white collar work but couldn't figure out how to use it right.
I feel like the authors make a logical inconsistency. They present the drop in "identify missing context" behavior in artifact conversations as potentially concerning, like people are thinking less critically. But their own data suggests a simpler explanation: artifact conversations show higher rates of upfront specification (clarifying goals +14.7pp, specifying format +14.5pp, providing examples +13.4pp). It's obvious that when you provide more context upfront, you end up with less missing context later. I'd be more sceptical about such research.
You could arrive at the essence of this by just having read and internalized Carl Sagan's The Demon-Haunted World. Especially the Baloney Detection Kit.
In my experience good prompting is mostly just good thinking.
And having the experience and judgment to ask the right thing.
And being willing to be wrong and to be misled; finding ways to contain that or build forcing functions against it.
In a strange way that's exciting, because it forces me to learn. And sometimes forces me to confront whether stuff I had was domain knowledge or portable as experience.
To the extent that this should be a thing, there are very few people I would want doing it less than the company who has repeatedly been caught lying about its product's achievements. Anthropic should not be taken seriously after their track record.
Honestly to use llms properly all you need to know is that it’s a next word (or action) prediction model and like all models increased entropy hurts it. Try to reduce entropy to get better results. Rest is just sugarcoated nonsense. To use llms properly you need a physics class.
And then some alignment, prompting structure, and task decomposition.
And praying that your desired output was embedded into the training data that was used to generate the model.
Which class? Or what subjects