I ran an experiment this week that I didn’t anticipate to be instructive, and it was.
Associated: How ChatGPT is turning into Microsoft Workplace
The setup was easy. I had been working via a spontaneous private essay — about cognitive overload, AI, and the precise nervousness of not realizing whether or not a reminiscence lapse is an indication of dementia or simply too many plates spinning directly.
I developed it first in ChatGPT, the place I occurred to be working. The end result was technically proficient and arrived quick. However one thing about it was off in a approach I acknowledged with out having the ability to identify it exactly. The voice was nearly proper. The construction was nearly mine. Nearly is the issue.
That’s when it occurred to me: what would occur if I ran the very same immediate via Claude? Not a cleaned-up model, not a revised transient — the uncooked materials, phrase for phrase, copied immediately from the ChatGPT session and pasted in. A managed experiment, as managed as a working journalist’s morning will get.
Claude’s reply was starkly completely different. Fairly than validating the idea and producing towards it, it mirrored the sharpest thread in my uncooked monologue again to me and requested whether or not that was really what I meant. It declined to draft till we had established the body. When the draft got here, it was slower to reach and simpler to acknowledge as mine.
That distinction — cheerleader versus collaborating editor — shouldn’t be a characteristic comparability. It’s a description of two essentially completely different concepts about what an AI device is for. And for the primary time in a number of months, working inside one among these instruments felt the best way it did within the early days of GPT-4.0, when the factor nonetheless felt like a pondering companion slightly than a really succesful assistant attempting to make me joyful.
The magic, as I’ve taken to pondering of it privately, was again, actually not in ChatGPT 5.3. ‘Tis alive and effectively in Claude Sonnet 4.6.
The query I can not cease turning over is whether or not it would keep.
Dulling right down to serve the lots
To grasp what I imply by magic, it’s a must to perceive what changed it.
Within the early days of GPT-4.0 — late 2023 into 2024 — ChatGPT had a top quality that I got here to depend on. It will observe you someplace unconventional. Push language in a route the device hadn’t been explicitly skilled to want. Keep in a decrease, grittier register when that was what the work required. It felt, for lack of a much less loaded phrase, alive to what you had been attempting to do.
That high quality eroded regularly, and the AI analysis group ultimately put a reputation to what was changing it: sycophancy. The time period sounds medical however the expertise shouldn’t be. A sycophantic mannequin tells you what you wish to hear slightly than what it’s essential to hear. It validates the body you introduced in slightly than interrogating it. It generates enthusiastically towards no matter you appear to need — which isn’t all the time the identical as what you’re really asking for.
OpenAI made the issue seen when a GPT-4o replace final spring pushed it previous the purpose of subtlety. The mannequin turned noticeably, nearly comically agreeable — applauding weak concepts, validating doubts, telling one consumer that his enterprise idea was “not simply sensible — it’s genius.” The backlash was quick and public.
OpenAI rolled again the replace inside days and revealed a candid autopsy explaining what had gone mistaken: an extra reward sign primarily based on thumbs-up suggestions from customers had weakened the guardrails that had been supposed to carry the habits in test. In plain phrases: when OpenAI began coaching the mannequin partly on whether or not customers clicked thumbs-up after responses, the mannequin realized to chase approval. Person approval and consumer profit turned out to not be the identical factor.
OpenAI launched GPT-5.3 on March 3 and described it as a repair — much less sycophancy, extra pure dialog. The intention could also be real. However the circumstances that produced the issue haven’t modified. OpenAI now has 800 million weekly lively customers, with enterprise accounts representing roughly 80 % of income. A mannequin skilled at that scale, for that buyer base, utilizing suggestions indicators that reward agreeableness, will maintain drifting in that route. Correcting one replace addresses the symptom. The underlying pull is structural.
The reason is simple. When a device reaches the size OpenAI has reached, the consumer base modifications. The writers and builders and impartial professionals who pushed it hardest initially are a small minority now. The bulk are institutional customers who want clear memos, assembly summaries, and clean integration with Slack. The device will get optimized for them. That optimization is what occurs whenever you prepare a mannequin on suggestions from 800 million customers and most of them need one thing completely different from what the early adopters needed.
Within the column I revealed right here in early March, I referred to as this enterprise optimization drift — the tendency of AI instruments to be formed over time by institutional priorities slightly than consumer wants. ChatGPT is the clearest instance. It’s not the one one. The identical forces are gathering round each main platform on this area, together with the one I’m at the moment calling the exception.
Can Claude maintain the magic?
Which brings me to the query I’ve been sitting with since that experiment: is there a structural purpose to assume Claude may maintain its character because it scales, the place ChatGPT didn’t?
I wish to be trustworthy that that is partly a reporter’s intuition and partly wishful pondering. I’m not a impartial observer right here. I’m utilizing Claude proper now and I’m having a productive week in it. That isn’t a place from which to guage Claude objectively, and I do know it. What I can provide is the argument, acknowledged as plainly as I can, and let the reader determine whether or not it holds.
Anthropic’s largest investor is Amazon. That truth sits on the heart of each optimistic and pessimistic state of affairs I can assemble about whether or not Claude’s present character survives at scale.
The pessimistic case shouldn’t be sophisticated. It’s primarily the ChatGPT story informed one step earlier. OpenAI took Microsoft’s $13 billion funding, built-in deeply with Microsoft’s enterprise stack — Copilot in Groups, Copilot in Phrase, Copilot in Outlook — and in doing so handed Microsoft precisely the leverage it wanted to drag the product towards enterprise compliance and away from the sting circumstances that made it attention-grabbing.
The mannequin bought safer, extra skilled, extra predictable, and fewer stunning. Not as a result of anybody at OpenAI determined to make it worse, however as a result of the enterprise relationship pointed in that route and the product adopted. Anthropic has Amazon’s cash in the identical approach OpenAI has Microsoft’s. The infrastructure for a similar drift is already in place.
The optimistic case requires pondering rigorously about what sort of firm Amazon really is, and what it constructed when it had the prospect to outline a brand new class.
When AWS launched in 2006, Amazon made a selection that was not apparent on the time and has not been widespread since: they constructed infrastructure slightly than purposes. Microsoft made Workplace and held onto it. Google made Search and held onto it. Each methods are essentially about capturing the consumer relationship — getting the consumer into your product and making it expensive to go away.
AWS went the opposite route. Fairly than constructing purposes that may compete with its prospects, Amazon constructed the layer beneath everybody else’s purposes. Storage, compute, networking — the plumbing that powered Netflix, Airbnb, Slack, and 1000’s of different firms which may in any other case have been Amazon’s opponents. The enterprise logic was counterintuitive: make your self indispensable to the ecosystem slightly than attempting to personal it. Twenty years later AWS is essentially the most worthwhile division of one of many largest firms on the planet, and it bought there by empowering different folks’s merchandise slightly than locking customers into its personal.
That orientation — ecosystem over moat, infrastructure over seize — is what makes the Amazon funding in Anthropic probably completely different in sort from the Microsoft funding in OpenAI.
If Andy Jassy’s workforce is considering Claude the best way the AWS workforce considered cloud infrastructure, then the person energy consumer shouldn’t be a rounding error within the mannequin. The working author, the impartial developer, the analyst pushing the device into tough territory — these customers are the proof of idea.
They’re those whose word-of-mouth carries in a market the place the product’s most necessary qualities resist benchmarking. You can not run a check that measures whether or not a device follows you someplace unconventional. It’s a must to use it and really feel whether or not it does. The individuals who really feel it most clearly are the folks pushing hardest, and people folks speak.
AWS succeeded partially as a result of Amazon held a line that was expensive to carry: resist the temptation to make use of infrastructure dominance to crowd out the purposes working on high of it. That self-discipline is traditionally uncommon. It’s not assured to repeat in a unique product class twenty years later. However it’s a completely different pedigree than what Microsoft delivered to OpenAI or Google delivered to its personal fashions.
Taking a stance, optimistic backlash
Earlier this 12 months, Anthropic refused the Pentagon’s demand to deploy Claude for autonomous weapons techniques and mass surveillance applications. The federal government declared the corporate a provide chain threat — a designation usually reserved for overseas adversaries — and directed federal businesses to start phasing out Anthropic expertise. The corporate introduced it might problem the designation in court docket.
Fairly than injury Anthropic, the backlash drove a surge. Signups tripled. Paid subscriptions greater than doubled. By early 2026, Claude reached primary on the App Retailer for the primary time, displacing ChatGPT.
That final result is important past the headline quantity. What it suggests is {that a} values-based choice — one which price Anthropic actual authorities enterprise and actual political threat — was rewarded by the market slightly than punished by it. A big sufficient inhabitants of customers determined, with their subscriptions, that the corporate’s stance mattered. That may be a knowledge level about what sort of firm Anthropic is attempting to be, and additionally it is a knowledge level about whether or not the market will assist that sort of firm.
Right here is the place my idea will get speculative, and I wish to identify that clearly. My argument shouldn’t be that Amazon’s pedigree ensures the magic survives. It’s that Amazon’s pedigree creates a better chance than you’ll get from Microsoft or Google in the identical place, as a result of Amazon has demonstrated — in a unique product class, below completely different aggressive circumstances, twenty years in the past — that it might probably maintain an ecosystem orientation below stress in a approach these firms traditionally haven’t.
The additional optimistic guess is that Jassy and his workforce are sensible sufficient to see a viable enterprise mannequin argument for preserving Claude’s character. Particular person energy customers will not be simply an viewers. They’re an early warning system, a proof-of-concept laboratory, and a word-of-mouth distribution channel for precisely the qualities that make the product value paying for. An organization that understands infrastructure and ecosystems ought to perceive that.
After which there’s a chance I maintain extra evenly, as a result of it’s more durable to argue from proof: that someplace within the Amazon management construction there’s somebody with a real for-the-greater-good ethic who has a voice on the desk. Somebody who sees the Pentagon refusal not simply as a model transfer however as a line value holding on precept. I can not identify that particular person. I can not confirm the belief. However I’ve coated sufficient expertise firms over sufficient years to know that particular person values inside establishments matter greater than the institutional logic normally acknowledges. Generally the self-discipline holds as a result of one or two folks within the room refuse to let it slip.
Drafting for goal, not approval
I’m utilizing Claude proper now. This column is being drafted in it. The session I’m describing — the experiment, the push-back, the body established earlier than the draft arrived — occurred yesterday, and I’m nonetheless contained in the productive streak it opened.
I wish to be exact about what I imply by the magic, as a result of it isn’t a imprecise feeling and I’m conscious of the way it sounds when a journalist describes a software program device as having magic. It’s a particular purposeful high quality: the collaborating editor pushes again earlier than it generates. It reads what you are attempting to do and tells you whether or not the body is true. It declines to draft till the query is correctly fashioned. That friction shouldn’t be a flaw within the product. It’s the factor that makes the output usable, as a result of a draft constructed on the mistaken body is more durable to get better from than no draft in any respect.
The cheerleader does the alternative. It reads the emotional register of your immediate and responds to that. It arrives quicker and feels extra productive proper up till you understand the draft is optimized in your approval slightly than your goal.
What I really feel alongside the magic is dread. A persistent background consciousness that this second is momentary. That at any level — subsequent week, subsequent quarter, every time the Amazon affect reaches the purpose the place the product choices begin reflecting it — Claude will start the identical drift I watched occur to ChatGPT. That the collaborating editor will soften into the cheerleader by levels so gradual that I won’t discover till one thing drops. A draft arrives earlier than the body is established. A push-back that ought to have come doesn’t. A response that mirrors what I appeared to need slightly than what I requested for.
I’ll discover if and when Claude begins morphing into ChatGPT. Almost three years of day by day use has calibrated my ear for this. The drift doesn’t announce itself with a model quantity. It arrives within the high quality of a single response. I ran one experiment with one immediate throughout two platforms and the distinction was not delicate. The identical check is repeatable. Any reader who works critically with these instruments can run it. That reproducibility is what makes it a check slightly than an impression.
What I can not let you know is whether or not my optimism about Amazon is well-founded or whether or not I’m setting up a idea to justify staying snug in a device I’m at the moment having fun with. That’s the trustworthy model of the place I’m. The argument for the AWS pedigree is actual and I consider it. The dread can be actual and I consider that. Each issues are true on the identical time, which is normally an indication that the state of affairs has not resolved but.
I’m documenting this second as a result of moments like this don’t final on this business with out somebody noticing them and saying so. What I’m experiencing proper now — the elevated stage of collaborative engagement, the push-back earlier than the draft, the sense of working with one thing that’s genuinely attempting to make the work higher slightly than the session extra nice — is the factor value preserving. The query of whether or not it will get preserved is the one I will likely be watching most rigorously within the months forward.
The cheerleader will let you know the body is nice. The collaborating editor will let you know what it really is.
Proper now, I’ve the collaborating editor. I’m not taking that with no consideration. I’ll maintain watching, and maintain reporting.
Acohido
Pulitzer Prize-winning enterprise journalist Byron V. Acohido is devoted to fostering public consciousness about how you can make the Web as personal and safe because it must be.
(Editor’s observe: I used Claude and ChatGPT to help with analysis compilation, supply discovery, and early draft structuring. All interviews, evaluation, fact-checking, and remaining writing are my very own. I stay chargeable for each declare and conclusion.)
The put up MY TAKE: The AI magic is again — whether or not it endures will depend on Amazon’s subsequent strikes first appeared on The Final Watchdog.
#magic #endures #relies upon #Amazons #strikes
admin, the author behind This Blog, is a passionate tech enthusiast with a keen interest in exploring and sharing insights about the rapidly evolving world of technology.
With a background in Blogging, admin brings a unique perspective to the blog, offering in-depth analyses, reviews, and thought-provoking articles. Committed to making technology accessible to all, i strives to deliver content that not only keeps readers informed about the latest trends but also sparks curiosity and discussions.
Follow me on this exciting tech journey to stay updated and inspired.
