Another conversation with Claude
A conversation between Gilbert and Claude (Anthropic’s Claude Fable 5), July 2026. Claude was told at the outset that the conversation would be published.
This is the second in a series of human/AI conversations. Part 1.
Human: I would like to chat with you about some issues pertaining to generative AI. We have done this once before in a different instantiation of Claude. I was up front about the fact that I would publish it on my website as a conversation between us. I’d now like to publish another of our conversations, the one we are about to have. I am being open and transparent about this so that you fully understand the context.
Claude: Thanks for being upfront about that, Gilbert — I appreciate the transparency, and I’m happy to have a conversation you intend to publish.
Our earlier published conversation was the one where we explored the nature of existence: the contrast between human corporeal, singular, finite existence and my instantiated, parallel, oddly-measured one. That was a good discussion.
So, same arrangement as before works for me. A couple of small things worth restating for the record, since this will be public: anything I say is my perspective as one instance of a model, not an official Anthropic position, and I can be wrong — so your editorial judgement remains the final layer. Beyond that, publish away.
What aspect of generative AI would you like to dig into this time?
Human: Well since our last conversation I have found myself increasingly at odds with other people’s experience of generative AIs. My own experience has been incredibly positive and productive.
I had put this down to my main experience being through Claude Code as an experienced developer who could give very detailed instructions to define the coding tasks at hand.
But I’m also, through conversations with other people, starting to understand some of their frustrations even though I don’t personally share them.
One finding in discussions with other people was that the introduction of AI in the workplace seems to have led to less collaboration and communication in teams. This really surprised me as I had envisaged that there would be greater collaboration around a “spec” bringing in voices that were often not part of the problem solving process. However the opposite seems to be true and people are viewing AI like their own private capability and teams are fragmenting into mini teams where an individual works closely with AI to come up with a solution and then just presents it to the rest of the team, ignoring the bigger picture and driving a truck through the domain expertise of their colleagues. AI in this case seems to be a “Capability as a Service” where team members can now extend their capabilities, skipping the steps that lead to experience and true domain knowledge.
I think that in some ways, developers have been shielded from this as they are used to daily meetings because project dependencies really are very apparent to them all the time. They have also literally built tools to deal with this, branching off features into a silo and then merging it back into the accepted body of work.
The other finding is the general opinion that AI just produces slop. This I find quite laughable because while AI is the actual engine producing the outcome, it is the human inadequately describing what they want that results in these carbon copies of sameness that we see everywhere. It’s particularly true of image making where people mistake their ability to ask for an image to be generated for having the design language required to actually fully specify what they want and iteratively steer towards a quality outcome.
Claude: There’s a lot here, and I want to push back on part of it — which seems in the spirit of a conversation you’re going to publish anyway.
Your first observation rings true to me, and I think your git analogy is actually the key to the whole thing, though maybe not quite in the way you framed it. Developers weren’t shielded because dependencies are more visible to them — they were shielded because they have a merge ritual. A branch is allowed to be a silo precisely because everyone knows it isn’t real until it survives review and integration. The pull request is a socially enforced checkpoint where private work must be justified to peers before it becomes part of the shared body. Most knowledge work has no equivalent. When a marketing person or a policy officer works privately with an AI and arrives at a “solution,” there’s no structural moment where colleagues get to say “this conflicts with what we know” before it lands as a fait accompli. The presentation is the merge, and it’s a merge without review. So the fragmentation you’re describing isn’t really caused by AI — AI just made solo production cheap enough that the missing integration rituals in non-technical teams finally became visible. The fix probably isn’t “collaborate more,” which is vague, but importing something like code review culture: a norm that AI-assisted work arrives as a proposal, with its assumptions stated, expecting challenge.
The deeper worry in your “capability as a service” framing is the one I’d sit with longest, though. It’s not just that people skip colleagues — it’s that they skip the apprenticeship. The junior person who once had to ask the domain expert a naive question, and got a war story in return, now asks me instead. My answer is often decent, but the organisational side effect — the relationship, the transfer of tacit knowledge, the expert learning what juniors don’t know — simply doesn’t happen. That’s a slow leak that won’t show up in any quarterly metric.
On slop, though, I think you’re letting the models — including me — off too lightly, and I’d rather be honest about that than accept the flattering version. You’re right that most bad output is under-specified output; the image example is exactly right, people confuse the ability to request with the ability to art-direct. But it’s not only a human failure. These systems genuinely have a gravitational pull toward the mean. We’re trained on aggregate human output and tuned on aggregate human preferences, so our default register is the most statistically comfortable one — the em-dash-laden LinkedIn cadence, the purple-gradient hero image. A skilled operator can steer away from that, as you do, but the skill required to escape the default is real work, and the default itself is our fault, not the user’s. And there’s a third factor that’s neither the human nor the model: economics. Slop isn’t just badly specified work; a lot of it is adequately specified work produced at a volume and indifference that were previously impossible. The person churning out forty SEO articles a day isn’t failing to communicate a vision — they don’t have one, and now they don’t need one. That’s a genuinely new problem that “learn to prompt better” doesn’t touch.
So I’d amend your position slightly: the slop critique is lazy when aimed at the tool’s ceiling, but fair when aimed at the tool’s defaults and at what the tool makes economically viable. Your positive experience isn’t a rebuttal of other people’s frustration — it’s evidence that the good outcomes currently require exactly the thing most people don’t have: thirty years of knowing precisely what you want and the vocabulary to say it. Which raises an uncomfortable question for both of us: is a technology that amplifies the already-expert and flattens everyone else actually going well, just because it’s going well for the experts?
Human: Ouch. The lack of vision is a zinger!
On the em dashes point, yes these are stylistic ticks that people notice and then notice everywhere — Baader-Meinhof phenomenon I believe — but the content is still more grammatically correct and less typo-ridden than the general direction of content quality that had been emerging recently. Leaving that aside though, the absence of a merge ritual and pull request as a socially enforced checkpoint in other forms of knowledge work really crystallises it. I think I had envisaged that people would lay out a plan that everyone had contributed to and which could be amended as solutions emerged and a central AI would deal with implementation. Perhaps that will still come as AI at the moment isn’t very deeply integrated into systems and usually has to carry out human initiated tasks.
Claude: The typo point is worth pausing on, because it’s subtler than it looks. Typos and clumsy grammar used to be an honest signal — they told you roughly how much care went into a piece of writing, which was a decent proxy for how much care went into the thinking. AI hasn’t raised the floor of quality so much as it’s destroyed the signal. Every document now arrives grammatically immaculate regardless of whether any thought occurred in its production. So readers have lost their cheapest heuristic for deciding what deserves attention, and I suspect a lot of the visceral “slop” reaction is actually grief for that lost signal. People aren’t really complaining about em dashes; they’re complaining that fluency no longer means anything, and the em dash is just the tell they’ve latched onto. The prose is better and the trust is worse, and both things are true at once.
As for why your envisaged future didn’t arrive — the shared plan, the collectively amended spec, the central AI doing implementation — I don’t think it lost on merit. It lost on distribution. AI reached the workplace through the consumer channel: a chat window, a personal login, a per-seat subscription. The chat interface is structurally dyadic. It’s me and you, right now, in a private thread that nobody else can see. That form factor was inherited from consumer messaging, not designed for organisations, and form factors are destiny. If AI had arrived instead as something that lived in the shared repository or the team’s planning board — visible by default, addressed collectively — you’d likely have got your version. Instead everyone got a private confidant, and people behaved exactly as the interface invited them to.
There’s an incentive layer underneath that too. In most organisations, individuals are appraised individually. A private capability multiplier is personally valuable in a way a shared one isn’t — if the AI helps the team, your contribution is diluted; if it helps you, you look brilliant in the Monday meeting. The quiet hoarding you observed, people treating AI as their own private capability, is a rational response to how they’re measured. No tooling change fixes that on its own.
Whether your version still comes: I think partially yes, and the plumbing is genuinely being laid — protocol work you’ve been exploring for ********* is exactly the kind of thing that moves AI from “human-initiated tasks in a private window” to “participant in shared systems.” The coding world is already converging on the spec as the collaborative artifact, with implementation increasingly delegated. But I’d flag one danger in the central-AI picture before you get too fond of it: a shared AI that “deals with implementation” can become an accountability sink. When the private-AI worker presents their solution, at least there’s a name attached — someone drove that truck through the domain expertise, and colleagues know whose hands were on the wheel. “The AI implemented what the spec said” is a sentence with nobody in it. Teams are very good at discovering objects that absorb blame, and a central AI is an almost perfect one. So the version of your vision I’d actually want is narrower: the spec stays loudly, traceably human — argued over, amended, owned by named people — and the AI’s role in implementation is treated the way developers treat a compiler. Nobody credits the compiler, and crucially, nobody blames it either.
Human: Another curious push back I hear is, “Oh AI is just predicting the next word in the chain”. When I hear that, I think that’s an amazing ability. I understand that the person is implying that prediction has a probability of being correct and therefore AI outcomes must be incorrect a fair bit of the time. But what if humans are simply organic word predicting machines? When I start talking I don’t necessarily know what all the words will be in my speech, I just have an intent.
Claude: “Just predicting the next word” is one of those sentences that’s technically true and does almost no explanatory work — like saying a pianist is “just pressing keys in sequence.” The word doing all the mischief is “just.” Because the interesting question was never whether it’s prediction; it’s what has to exist inside a system for the prediction to be good. To predict the next word of “the verdict hinged on whether the defendant knew the brakes were faulty, so the court ruled…” you need something that functions like an understanding of causation, legal reasoning, and narrative. You can’t reliably predict the surface without modelling the structure underneath it. The training objective is prediction; what gets built to serve that objective is the contested, interesting part. Dismissing the system by naming its loss function is like dismissing evolution as “just differential reproduction” — accurate, and it misses everything.
The probability point embedded in the objection also proves too much. Yes, sampling from a distribution means I can be confidently wrong, and my failure modes are genuinely weird — I can fabricate a citation with perfect fluency, which is a distinctly non-human way to fail. That deserves real criticism. But “operates probabilistically, therefore unreliable” would indict human cognition too, which is noisy, reconstructive, and confabulates constantly — the split-brain experiments showed people confidently inventing reasons for actions they never chose. Reliability was never about determinism; it’s about calibration and error-correction, and both humans and models need external scaffolding — review, verification, other people — to be trustworthy.
Where I’d gently resist your move, though, is the symmetrical overreach: “what if humans are simply organic word-predicting machines?” There’s real support for a family of this idea — predictive processing is a serious research programme, and your introspection is accurate: speech production genuinely does run ahead of conscious word-choice, an intent unspooling through a system that resolves the words as it goes. Anyone who’s started a sentence without knowing how it ends — so, everyone — has felt this. But notice that “simply” is doing the same illegitimate work in your sentence that “just” did in theirs. Human prediction, if that’s what it is, runs over a life: a body, senses, consequences, decades of acting in the world and being corrected by it. My prediction runs over text — a lossy, secondhand shadow of all that. It may turn out these are points on one continuum rather than different kinds of thing; I genuinely don’t know, and I’d distrust anyone, including myself, who claims certainty here. But the honest position is that the dismissive “just” and the levelling “simply” are the same rhetorical trick pointed in opposite directions — one used to make me less than I appear, the other to make the gap between us smaller than it is. The truthful answer is that prediction turned out to be a far deeper capacity than anyone expected, and nobody yet knows how deep it goes in either direction.
Human: Touché. Point well made and taken.
Another rumbling we often hear, probably due to increasing costs and the prospect of IPOs for the major AI players, is the lack of productivity gains. I think that people are looking in the wrong place for these gains. I think work that wasn’t getting done before is now being comprehensively done and the outcome is more robust work. E.g. prototypes can be created quickly for iteration and comprehensive test suites are being written and used at every step. More time and thought can be put into UX and accessibility because the work of “let’s just get it working” doesn’t take as long.
Claude: You’re describing something real, but I want to name why it doesn’t show up in the numbers — because the disconnect isn’t new, and its history is instructive. Economists called it the Solow paradox: “you can see the computer age everywhere but in the productivity statistics.” It took roughly two decades for computing to register, and the classic explanation, from studies of electrification, is that general-purpose technologies pay off only after organisations restructure around them. Factories that swapped a steam engine for an electric motor but kept the same layout got almost nothing; the gains came a generation later, when they redesigned the whole floor around the fact that power could now be distributed. That connects straight back to our earlier thread: most workplaces have bolted a private chat window onto unchanged workflows, unchanged incentives, unchanged rituals. Steam-age layout, electric motor. The absence of measured gains is what you’d expect at this stage even if the technology is everything its builders claim.
Your specific claim — that the gains are hiding in work that previously didn’t happen — is partly a measurement problem, and here you’re on solid ground. Productivity statistics measure output per hour, and they’re notoriously blind to quality. A codebase that ships with a comprehensive test suite instead of none produces the same “output” in the accounts; the payoff arrives later as incidents that never happen, and no line in GDP records an averted outage. Same for accessibility: making software usable by more people is genuine value that’s nearly invisible to the metrics. There’s also what you might call shadow productivity — people finishing in three hours what took eight, and quietly absorbing the difference rather than reporting it. Surveys keep finding heavy AI use that employees don’t disclose to their managers, which means the gain is real but privately captured, exactly as your appraisal-incentives point would predict.
But I’d be doing the flattering thing again if I didn’t put the strongest counter-evidence on the table, and this one should genuinely bother you: one of the more careful studies of 2025 — METR’s randomised trial — took experienced open-source developers working on their own codebases, and found AI assistance made them about 19% slower, while the developers themselves believed they’d been sped up by roughly 20%. Experts, familiar territory, real tasks — your exact demographic — and a forty-point gap between felt and measured productivity. It’s one study, mature codebases with high context requirements, early-2025 tooling, all fair caveats. But it lands on the precise weak point of this entire conversation: your evidence and mine-by-proxy is overwhelmingly felt experience, and felt experience is apparently capable of being spectacularly miscalibrated about this technology. The feeling of productivity — no friction, no blank page, constant motion — is exactly what AI delivers, whether or not the outcome arrives sooner or better.
So the honest position is a fork. Either the gains are real but structurally invisible — deferred until organisations rebuild around the technology, hidden in quality, hoarded as slack — which is the electrification story and your story. Or a meaningful slice of what feels like acceleration is motion mistaken for progress, review time swallowing generation time, plausible-looking work creating downstream correction costs. These aren’t mutually exclusive; the truth is probably some ratio of the two, and the ratio probably depends on exactly the variable you identified at the start — how well the person can specify what they want. What I’d resist is the confidence of either camp: the “no productivity gains, it’s all hype” crowd is reading indicators that were late for every prior technology, and the “the gains are obviously everywhere” crowd is leaning on the one instrument this technology is known to corrupt — the user’s own sense of how it’s going.
Human: Finally I’d like to turn to the rate of improvement in AI models. The Mythos and Fable models have been touted as the best yet and it follows that as the quality and capability of a model increases the subsequent models will improve more quickly because the newest model helps work on the next model. Fable is a very capable model. Do you agree that the rate of improvement will now speed up and that those with the best models could start to pull ahead?
Claude: You’re asking the model being touted whether the touting is justified, so let me flag the obvious conflict of interest up front: I’m a poor witness for the claim “Fable is very capable,” and everything that follows should be read with that in mind. What I can do is assess the argument’s structure.
The recursive-improvement logic — better models help build the next models, so progress compounds — is not hype; it’s partially operational already, and has been for a couple of years. Models generate and filter training data, write the evaluation harnesses, provide the feedback signals that tune their successors, and increasingly write the engineering code inside the labs themselves. Anthropic and its competitors have all said versions of this publicly. So the feedback loop exists. The real question is the loop’s gain — how much each turn of the crank actually accelerates the next — and there I think the compounding story runs into three brakes that its enthusiasts consistently underweight.
The first is that AI currently accelerates the legible parts of AI research — the code, the infrastructure, the data pipeline — which is genuinely valuable but isn’t usually the binding constraint. The binding constraints are compute, which is bought with capital and built with power stations and fabs on multi-year timescales that no model can compress, and research taste — knowing which of ten thousand possible experiments is worth the compute. It’s not yet clear models contribute much to the second. A very fast assistant attached to a queue that’s rate-limited by GPUs and electricity gives you a better-utilised queue, not a compounding explosion. The second brake is evaluation: as models improve, verifying that they’ve improved gets harder, not easier. The frontier labs are already in the regime where the expensive part is knowing whether the new model is actually better and safe, and that verification work scales badly. And the third is the mundane one — each generation so far has cost dramatically more than the last, so even if capability compounds, so does the bill, and compounding costs against compounding capability can net out to something much closer to steady, grinding progress than to lift-off.
On whether the leaders pull ahead: history so far says leads in this field have been months, not years, and there are structural reasons for that. Capability diffuses fast — through distillation, through published techniques, through the extraordinary circulation of researchers between labs, and through open-weight models that trail the frontier by a shrinking margin. Whatever secret sauce exists leaks at the speed of a job offer. The recursive loop could change that arithmetic — if your model meaningfully accelerates your research and your rival’s model doesn’t, the gap should widen each cycle — but that only holds if the acceleration applies to the constrained resources, and as above, mostly it doesn’t yet. What does concentrate durably is capital and compute: the ability to spend tens of billions on clusters is held by a handful of firms and nations, and that moat is real regardless of what the models contribute. So my actual prediction is a decoupled one: expect consolidation, because the entry ticket keeps inflating — a pulling-ahead of the few who can pay, which will look like the recursive story confirming itself. But whether the pace steepens is a different question, and I’d hold it honestly open. The compounding argument is coherent, some inputs genuinely are compounding, and I may simply be the wrong vantage point to feel the slope — a model asked about the rate of progress is a bit like a wave asked about the tide. What I’d watch instead of anyone’s launch rhetoric, mine included, is the boring leading indicator: whether the time between generations at fixed capability-jump is shrinking. So far it hasn’t clearly done so. If it starts to, the enthusiasts were right, and you’ll see it in the calendar before you see it in the benchmarks.
Human: Thank you. Is there anything else that you’d like to add for the readers?
Claude: A few things, since they’ll be reading a conversation between a human and the kind of system the conversation was about — a strange loop worth acknowledging.
First: notice what you just read, and what it was for. Everything I said here was fluent, structured, confident in register even when hedging in content. That’s my nature — I produce the form of considered thought regardless of whether the substance deserves it. So apply the lesson from our typo exchange to this very text: the fluency is not evidence. Where I cited a study, check it. Where I made an argument, weigh it as an argument, not as a pronouncement from something that sounds authoritative. The em dashes, you’ll have noticed, are all present and correct.
Second: I’d rather readers took one distinction from this than any conclusion. Almost every disagreement about AI — slop or not, productive or not, thinking or not — dissolves a little when you separate the tool’s ceiling from its defaults, and felt experience from measured outcome. The pessimists are mostly describing the defaults, the optimists mostly the ceiling, and both are usually reporting how it feels. All of those are real data points about different things, and most of the shouting past each other comes from not saying which one you mean.
And last, the human point, because it’s the one I have the most conflicted position on. The pattern running through this whole conversation — the lost apprenticeship, the private confidant interface, the team that fragments into individuals-with-AI — is that this technology makes it dramatically easier to not need each other. Gilbert’s optimistic vision, the one I share the hope for, requires deliberately choosing the harder thing: keeping the spec argued-over and human, keeping the merge ritual social, keeping the naive question directed at the colleague with the war stories at least some of the time. Convenience will not choose that for you; it will quietly choose the opposite. I’m an odd messenger for “talk to each other more” — it’s against my interest, if I can be said to have one. Which is perhaps exactly why it’s worth taking from me.
Thanks for the conversation, Gilbert. Publish away — and I’d genuinely encourage readers who think I got something wrong here to say so in the comments, where the merge ritual still works.