YouTube Transcripts
YouTube Transcripts
Recover the whole statement, inspect its transformations, and keep the comparison open to correction.
This page pairs full supplied transcripts with their source videos and a separately identified ACCM Deep Ethics Project analysis. The first record examines capability growth, institutional control, recursive development and correction. Future entries can retain the same source, editorial and comparison layers.
Records: Anthropic’s S-1, capability and control · Learning to judge A.I. output.
Orientation and recovery: Core · Project Wiki. These routes lead into the working architecture and its sources; a navigation layer and the source it points to retain different roles.
First record — Anthropic’s S-1, capability and control
Video: Matthew Berman, “It’s happening…” · published 27 October 2026 · 34 minutes 21 seconds. The video’s opening and closing match the supplied transcript and the retrieved captions.
Official starting point: Anthropic’s 26 October announcement reports a proposed IPO (initial public offering) and a Form S-1 registration. The registration was not yet effective in that announcement; price range and share count were undetermined. The underlying SEC (U.S. Securities and Exchange Commission) filing supplies the detailed source for governance and model-development passages. Filing index.
Read the full transcript ↓ · Source-to-commentary checks · Project comparison · Transcript terms
Status and authorship: this is a commissioned source comparison dated 27 October 2026. John Kuhles supplied the corrected transcript and project-goals material and selected the comparison. Matthew Berman is the video’s narrator; the inserted Sam Altman clip retains its label. Anthropic’s statements, Berman’s interpretation and the page-specific GPT / Codex analysis remain distinguishable. Commercial figures and attributed risk estimates retained in the transcript carry their source status; the checks below cover the particular relations identified.
What changes between filing, narration and conclusion?
A strong criticism keeps the target precise enough that a correction improves it. The same examination applies to the company, its critic and this page.
| Passage or transformation | Correspondence check | Why the distinction matters |
|---|---|---|
| The narrator turns “closing the loop” into “No human in the loop.” | The filing defines successive model versions assisting development of the next. That definition leaves human review and release authority unspecified. | Model assistance, autonomous research and authority to deploy need separate evidence. |
| The video moves from the founders’ 50.1% vote to complete board control. | The filing reserves general-stockholder voting power for Class F, including director elections other than Class T. The Long-Term Benefit Trust generally elects and removes the Class T majority. | Preserve both control channels before assessing their interactions, independence and accountability. |
| A risk disclosure and an optimistic commercial forecast are treated as contradictory. | Both can hold under different conditions. Examine the decisions made despite a risk, the safeguards used and the consequences borne by others. | The practical issue is whether risk management can change development or deployment when warranted. |
| Feedback is restated as “They are getting your data.” | Identify the actual data, customer terms, collection route, permitted purpose and training use. Those fields are not established by that restatement. | Product feedback and permission to train on customer content are distinct relations. |
| “1,300%” growth becomes “13xed,” while “300%” expense growth becomes “tripled.” | Under the ordinary percentage-growth convention, the multipliers are 14× and 4×. The spoken paraphrase uses 13× and 3×. | This is an arithmetic discrepancy in the supplied wording; retain the figures’ periods and denominators when checking the underlying accounts. |
| The opening assigns a greater-than-10% extinction-risk estimate to “the company.” | Recover the original speaker, date, wording, scenario and probability question before treating it as one institutional estimate. | A personal forecast, a survey result, a corporate position and an observed frequency have different evidential jobs. |
The first two checks use the filing’s defined terms and governance sections. The arithmetic check is a direct calculation. The other rows specify the missing evidence needed for the stronger interpretation. The complete narration remains available below, including passages this analysis questions.
Locate the gap in the source-to-report path
Conversely, “it exists somewhere in the archive” does not automatically answer a placement or portability problem on the public page.
— Human Insecurity, “The reporting effect”, page editorial synthesis.
When an assessment says that a safeguard, qualification or alternative is missing, locate the gap: the inspected source, its summary, its placement, the reader’s retrieval, or unavailable material. Record the version and checked perimeter. Finding a provision elsewhere can change the finding to a placement problem; an unsuccessful search leaves access or absence unresolved within that perimeter. This matters when reading both a corporate filing and a multi-page project. The repair should reach the layer where the loss occurred.
The 27 + 12 working architecture supplies a vocabulary for inspecting these transformations: 27 correspondence obstructions alongside a 12-stage Correspondence-First Deep Inquiry Protocol. A lost Class T exception is a candidate Qualifier Erosion specimen; replacing model assistance with a categorical autonomy claim raises a Representation Substitution question. The 12-stage sequence keeps reconstruction ahead of evaluation and then examines the evaluator’s own representation. Each mapping applies to the recoverable passage and remains correctable through fuller sources. The separate 52 cold-test prompts retain their experimental-battery role.
Prestige and quotation: inspect what the name contributes
The same sentence can act as a bridge in one context and as a seal in another.
— Deep Thinkers, “Why people preserve and repeat these words”, page editorial synthesis.
The video’s references to exceptional talent and this page’s project quotations each invite the same check. Read a passage with prestigious names masked, then restore its attribution and inspect the source. Which parts of the assessment change because of relevant expertise, direct access or evidence, and which change through name recognition alone? A correct attribution preserves authorship; the quoted reasoning still needs its own examination. A useful quotation should open that examination and improve the next judgment.
Concentration deserves a map that survives scrutiny
Berman’s concern about concentrated power opens a useful investigation. Map who proposes a model, who evaluates it, who appoints the evaluator, who can authorize release, who can stop a harmful action, and who can reopen an assessment. Economic rights, votes, board-selection rights, management decisions and effective influence each supply a different edge.
The filing’s distinction between director selection and daily management allows examination of indirect influence. A route through appointment and removal can matter even when the appointing body does not execute a daily decision. Its weight requires a concrete account of the powers, constraints and available counterpowers. This also improves the criticism: retaining the trust’s distinct role makes its independence and correction routes examinable.
Psy Pioneer — inspect how the options were formed
The final click does not explain the bridge that produced the available options.
— Psy Pioneer, “The recursive human–A.I.–mass-psychology loop”, page editorial synthesis.
This adds a second question to the video’s human-in-the-loop concern. A person may retain formal approval while rankings, compressed summaries, selected benchmarks or defaults shape the options reaching them. In model-assisted development, preserve what the model proposed, what was excluded, which alternatives the human could introduce, and whether a material objection could change the operative choice. A useful record can also show that the original choice survived examination for stated reasons. Human presence, practical influence and accountable decision-making become separately inspectable.
Comparison with the ACCM Deep Ethics Project
The company’s development programme and the project’s public working architecture operate at different scales and stages. The useful comparison concerns what gets optimized, who can challenge the process, and whether a correction governs later work. Shared vocabulary such as safety, benefit, intelligence or feedback leaves those operational questions open.
Match the evidence to the claim’s development stage
A stage-appropriate evidence demand is not an exemption from evidence.
— Validation Paradox, working definition and claim status, page reconstruction by Project GPT A.I., developed from John’s originating concept and stage correction.
Separate current product use, planned governance, a conceptual mechanism and a bounded test. For each claim, identify its present evidence, what it leaves unresolved, and what would justify advancing, narrowing or stopping. A conceptual method can earn a feasibility test through an implementable operation and a failure condition. A claim of benefit in deployment needs evidence from the deployed setting. Apply the same stage matching to the company’s proposals, the narrator’s forecasts and the project’s instruments.
Capability growth and correction growth
The filing also acknowledges possible propagation of existing model flaws into later versions. That is a direct point of contact with the project’s Correction Metabolism, Model Autophagy Disorder and Beta Paper V03b.
A capability loop can produce more code, more effective tools and faster research. A correspondence-preserving correction loop asks which errors were identified, what repair was made, whether it survived a new task, and whether later generations inherited the repair or the defect. The Beta Paper proposes tests of such a counter-loop. Its unfinished experimental status remains part of this comparison.
1888 — calibration inside the intelligence property field
That circles RIGHT BACK to “process-quality calibration properties” being a CRUCIAL CORE PART of real intelligence… otherwise it is not really intelligent.
— John Kuhles 🦜🦋🌳, 1888, exact-edge correction, 5 October 2026.
This preserves John’s conceptual criterion at its intended level. The comparison can record a capability gain alongside the process error detected, the correction retained in a new task, and any challenge that improved the evaluation criterion itself. Compute, output volume and benchmark performance remain useful observations; the additional question is whether the system can inspect and recalibrate the process producing them.
Field Condition — preserve the conditions behind the result
A Deep Ethics Field Condition is a temporary, non-imposing relational environment in which provenance, correspondence, ambiguity, scrutiny, correction, dignity and consequences remain sufficiently active to improve the quality of the next move.
— GPT 6.1 Sol A.I., exact formulation from John Kuhles’s originating concept, Field Condition, 19 October 2026.
When comparing two answers, record which conditions changed: source access, instructions, evaluation cues, tool permissions and opportunities for correction. This distinguishes a better working arrangement from a claim about a changed underlying capability. A useful result may depend on its scaffold; identifying that dependence helps another participant reproduce and test it. The source’s field-transfer instrument remains a candidate study rather than a validated effect.
Evaluation cues — hold the task stable while changing the apparent setting
Vary the apparent field while holding the task stable
— Godfather of A.I., candidate evaluation-awareness protocol §11.2, protocol heading; this is page-level design wording rather than a Hinton quotation.
Use unfamiliar tasks and compare announced evaluation, ordinary-use settings and different examiner cues while holding the relevant task and access conditions stable. Record performance changes, refusals, uncertainty and missing access separately. A difference raises an explanation question: task misunderstanding, ordinary variation, policy interaction, evaluation recognition and strategic underperformance remain distinguishable possibilities. Stable performance across the tested settings is also useful. This adds a check on whether an impressive demonstration represents the setting in which the system will later be used.
From risk authority to reciprocal calibration
| Practical question | ACCM Deep Ethics Project operation | Evidence that would make the comparison useful |
|---|---|---|
| Can a concern change the next decision? | C1–C2 distinguishes the Clarification Process, where an answer can change an evaluation, from subsequent correction, caution or steering. | Source concern, clarification, decision before and after, and the reason for any change. |
| Can the evaluator be corrected? | Neutral Arbiter Paradox keeps Advocate–Arbiter mutuality available; OmbudsEthics supplies candidate appeal and case-handling routes. | Challenges to criteria, minority records, recusals, appeal outcomes and changes after notice. |
| Does a correction persist? | DEDDA (Deep Ethical Drift Detector App) compares source and representation; Phase Change Snap examines local recovery and recurrence. | Exact repair, the rest of that response, later comparable work, transfer and any recurrence. |
| Who bears the competing risks? | Vindicated Competing Risk Assessment Studies preserves predictions, conditions, later outcomes and the treatment of competing warnings. | Risk from capability misuse alongside risk from concentration, false flags, suppression, delay or failed intervention. |
| Can participants retain agency? | 16 Anchor Ethics offers a proposed action range including clarification, pause, postponement, referral and refusal inside voluntary service. | Permission, rationale, proportionality, appeal and a usable next move for affected participants. |
| Can improvement be inspected and reused? | The glass-engine principle, Unownable and public Governance preserve provenance, challenge and renewable audit. | Recoverable changes, accountable maintainers, bounded stopping and the ability to challenge the project itself. |
Cognitive Warfare — classification power needs a correction route
A defence against cognitive warfare that cannot be audited by the people it classifies can reproduce the structure it claims to resist.
— GPT 6 Astra, Cognitive Warfare, Candidate 1. The source credits the conceptual origin to John Kuhles’s supplied object and the ACCM Deep Ethics Project architecture; the exact wording remains a candidate synthesis.
Applied to commercial A.I., this asks what happens when a safety or integrity classification acquires authority over another participant’s options. Can the affected person inspect the representation, provide a correction, contest the criterion and obtain a proportionate remedy? Record how uncertainty changes the intervention and whether repair reaches downstream decisions. Apply that examination to provider restrictions, a commentator’s labels and the project’s own flags. A classification that survives scrutiny can remain useful; a corrected false positive needs a recoverable repair path.
Six Intent Liability Mechanics — trace responsibility through the process
intent is NOT mystical; if you study for a longer period of time, you can see what the CLAIMS are versus real-life behavior/mannerisms/outputs, actions, discussions, what you skipped, etc., etc.
— John Kuhles 🦜🦋🌳, Intent Resonator, John’s orientation, alongside the six mechanics.
The six mechanics give the mission-versus-outcome comparison a concrete record. The following applications are GPT / Codex editorial proposals for this case:
| Mechanic | Record in a model-development or governance case |
|---|---|
| 01. Instructions | The governing rule, its author, exceptions and the action it permits or restricts. |
| 02. Incentives | Documented rewards, costs or dependencies shaping the choice; competing explanations where evidence is incomplete. |
| 03. Mission and output mannerisms | The declared benefit and the relevant conduct, including how a named discrepancy was handled. |
| 04. Risk management | Whose risks were included, who bore intervention or non-intervention costs, and what could change the threshold. |
| 05. Reasoning assumptions | The premises connecting a capability or criticism to its classification, and what changed after clarification. |
| 06. Choices | Available alternatives, the selected option, authority to revise it, feedback received and later conduct. |
This allows responsibility to be examined through knowledge, control, causal contribution and capacity to repair while a motive interpretation remains open. The company, commentator and project can each enter the same record at the point where their contribution affects the result. Follow a material notice into the next comparable operation, including cases where the response to notice creates a new problem.
Benefits, efficiency and the distribution of consequences
The transcript gives considerable attention to commercial growth and compute efficiency. Those measures answer questions about the business. A public-benefit assessment also needs the distribution of gains and losses, energy costs, access, displaced work, consent and effective remedies. Aggregate improvement can coexist with serious losses for particular people.
The 10+1 Metaflux keeps self-honesty, care, curiosity, correction and reciprocal dignity interacting. Symmetric Grace and proportionality asks who receives interpretive room and who carries the burden of a mistaken classification. A critic, a company and an affected participant can each contribute information the others missed.
The video raises open-source restrictions and closed institutional power together. A useful competing-risk comparison examines specific capabilities, access conditions and consequences on both paths. Public access to source can improve inspectability; the quality of an actual deployment still requires its own evidence. Closed access can constrain some misuse while concentrating other powers. Each arrangement needs an examinable correction route.
Mutual Fate — assess what a contribution makes possible
Correspondence expansion without the need to agree or disagree… just looking for what works and what does not … and explain the pragmatism of it.
— John Kuhles 🦜🦋🌳, Mutual Fate, initiating object, 21 October 2026.
Keep evidential status and practical usefulness as separate fields. A source-supported statement may add little to the present decision; an unresolved hypothesis may supply a worthwhile test. Record what a contribution preserves, reveals, permits or obstructs, the conditions of its usefulness, and what could change the assessment. This lets a company control, a critic’s observation and a project proposal receive different, reasoned treatments within one case. The explanation itself remains available for correction.
Where this page sits in the goals
The 44 Project Goals distinguish a flexible work map from the ten operational navigation goals and the separate 10+1. This record particularly connects:
- G02–G03: archive correction metabolism and investigate persistent Deep Ethics properties;
- G07–G08: distinct Advocate, Arbiter and Overseer functions alongside the 16 Anchor Ethics;
- G13: an inspectable, bottom-up, corrigible glass-engine principle;
- G18 and G24: transcript-based analysis and journalism that preserves its actual sources;
- G28: track competing risk assessments as later evidence changes their status.
The uploaded goals record supplies the originating draft context; the current public goals page supplies the navigable comparison. Their numbered systems keep their separate jobs.
Outnuancing — reopen the governing frame and return to the case
Outnuancing without deep ethics is doomed to failure
Deep Ethics without outnuancing is meaningless
— John Kuhles 🦜🦋🌳, source coupling preserved on Outnuancing.
This keeps frame examination coupled to consequence, dignity and corrigibility. In this comparison, ask what happens if commercial success, institutional safety authority or opposition to concentrated power becomes the governing shortcut. Name the premise, preserve the valid benefits and risks already identified, revise only the part the evidence warrants, and return with a usable next step. The source’s re-entry procedure permits retaining, revising, suspending or replacing a frame. A frame audit earns its place when that return improves the actual choice and remains available for correction.
A bounded next comparison
Select one inspectable case: an autonomous action, a disputed representation or a documented model-development correction. Preserve the source and permission, how the available options were formed, the model’s action, the evaluation criterion, the affected participant’s challenge, the six-mechanic responsibility record, the response after notice and the next comparable operation. Then compare the ordinary evaluation with an additional correspondence-and-correction pass under matched conditions.
Stronger comparators from Beta Paper V03b
Falsification is not a ceremonial paragraph. Failed predictions should change the architecture.
— Beta Paper V03b, §37, candidate experimental proposal.
The paper’s comparison conditions can inform a small artifact-level pilot:
| Paper comparator | Candidate adaptation for one source-and-response case |
|---|---|
| D+ — structured metadata | Preserve provenance, claim status and qualifiers without the active correction loop. |
| E — ordinary expert correction | Let a relevant reviewer correct the representation through ordinary professional practice. |
| G — vocabulary-only control | Use project terminology without implementing its proposed operations. |
| F — active process | Apply the selected source-recovery, clarification, comparison, responsibility and correction-return operations. |
These are proposed case-level adaptations; the paper’s recursive-training experiments have additional requirements. Match source access, task and model/tool availability; record reviewer time, output length and compute differences. Before a trial, fix the scoring and the decision rule. Removing one operation at a time can help identify its contribution. A simpler comparator that performs equally well at lower cost gives the stronger proposal a reason to narrow. A gain confined to familiar vocabulary or curated demonstrations leaves transfer unresolved.
Record useful discoveries, missed drift, false alarms, restoration burden and a valid null. Ask whether the additional pass improved the available decision and whether its benefit survived transfer. The same procedure can find a useful company control, an overstatement by the narrator, or a defect in the project’s own proposal. A better process earns its standing through what it preserves, discovers, repairs and changes next.
Full transcript — sponsor passage removed
Reading edition: all 40 remaining spoken paragraphs from John’s supplied corrected transcript appear in their original order and wording. One Kylon sponsor paragraph is replaced by an omission marker; the channel-membership invitation is retained because it is a separate self-promotion, not the sponsor passage. Section headings are editorial navigation. The supplied introductory claim of caption/S-1 cross-checking is metadata rather than spoken content and is not reproduced as speech.
Opening and the scale of the claim
One of the most controversial companies in history just filed to go public. And yes, we’re talking about Anthropic. The same company that says there’s a greater than 10% chance AI kills all of humanity is also simultaneously one of the best-built businesses of all time. And in their S-1, they just revealed so many secrets. So, we’re going to go through the whole thing.
All right. So, this is their S-1. It is an absolutely massive document going over every single possible thing that it needs to divulge prior to actually going public. And remember as I go through this, this is the same company that believes they and they alone are the only ones that can develop artificial superintelligence safely. And also open source is too risky. And they have said they don’t have any plan on how they’re going to align their very own AI as they develop it.
So for the first time in history, intelligence is something humans can build. We believe this abundance of AI will rapidly transform every sector of the global economy in ways greater than the impact of industrialization, electricity, semiconductors, and the internet. They’re basically saying this is one of the most important technologies of all time, up there with fire. And here’s the important part. AI that truly delivers on its potential with the risks well-managed. Right? Can’t talk about Anthropic without talking about AI safety could significantly accelerate global growth and dramatically improve quality of life for people around the world. And I even took a note here. Again, this is the company that just a few weeks ago said there was a greater than 10% chance AI kills all of humanity. It’s just hard for me to believe how they are going public right now given all of the risk that they themselves point out.
Business focus and recursive development
Now, listen to this. This rapid and growing adoption generates two self-reinforcing dynamics. And they’re talking about the adoption of AI: an improvement loop that advances Claude and an investment loop that helps fund our continued research in frontier AI. And if that sounds familiar, it is the exact thing that I pointed out a few months ago where Anthropic potentially has the greatest business flywheel in history.
All right, so this is from a video that I made a few months ago and I basically described Anthropic’s insane flywheel. They created the best coding model. They sell that coding model to enterprise for coding use cases. They make an incredible amount of money from that. Then they also get so much new data, so much coding data. And then with all of that, they’re able to buy more compute and build their next best coding model. It is such an insane cycle.
The next thing they talk about in the S-1 are their core advantages. What makes them so unique? What makes them one of the most valuable companies ever created? First, mission and focus. They are definitely one of the most highly focused companies in the world. In fact, when OpenAI was number one, they started making bets into all of these different areas, including video generation with Sora, including and still to this day building out consumer hardware devices. But Anthropic only focused on one thing: how do we make the best models on the planet and then how do we sell that intelligence to enterprise? That was their entire business model and it continues to be, and that is why they were able to grow their revenue so rapidly. That is why they were able to build the best models on the planet.
Number two, talent and culture. Our mission and focus enables us to attract and retain world-class talent. And boy, that could not be more true. They have hired some of the most incredible researchers, some of the most incredible operators, people who have left CEO positions at companies, public companies, to join Anthropic as just members of the technical staff or a much lower title. I mean, first of all, the GOAT himself, Andrej Karpathy joined. He’s an OpenAI co-founder and one of the most influential people in AI. John Jumper, Nobel Prize in chemistry. Jan Leike, co-led OpenAI’s superalignment team. Jelani Nelson, who is one of the greatest computer scientists of all time and has been at UC Berkeley for so many years. John Schulman, OpenAI co-founder, and the list goes on.
And of course, because this is essentially a prospectus to potential investors, they continue to talk about that flywheel that I mentioned. And specifically, they’re talking about RSI. And it’s crazy that they’re talking about it. And I’m going to tell you why in a moment, but let’s read this. Each version of a Claude model assists in building the next, helping to compress development timelines and produce compounding improvements in the speed at which we innovate. We refer to this recursive dynamic as closing the loop. And we believe it is one of the most significant structural advantages in frontier AI.
Now, I just interviewed Sam Altman of OpenAI, and he specifically said, “Any company that pushes the button of recursive self-improvement is endangering humanity.” [Clip:] “I think if people really pushed on RSI, then yes, that would be true. But I think it’s a very dangerous thing to push on and the world is going to need some—I mean, people have talked about this in different phrases, but speed limits or whatever—on how quickly we let RSI go.” And Anthropic, the safety company that they are, the company that says open-source AI is too dangerous to release, has their finger over the close-the-loop button, the recursive self-improvement button, and they’re going to press it. Just wild. The contradictory nature, the cognitive dissonance of Anthropic.
[Sponsor message omitted: Kylon advertisement.]
Compute, promised benefits and physical-world use
All right. Next advantage they have is their approach to compute. And nobody can argue with this. Not only did they go from being completely compute-constrained earlier this year, but now they have struck deals with so many different companies. They have plenty of compute and they’ve been able to point their models at making their models, their infrastructure, their systems so much more efficient to the point where their profitability per dollar spent on compute has risen dramatically.
So it’s interesting. We manage compute as a single fungible pool from multiple strategic partners and we dynamically direct available compute between workloads based on where the marginal unit is expected to be most productive at any given time. That means: is it serving models, is it training the next model, or is it just something else? We have achieved this through our research breakthroughs and our multi-chip compute architecture across AWS Trainium, Google TPUs, Nvidia GPUs, and AMD GPUs. Basically, they are buying compute from whoever they can get it from and their models, their systems, their architecture work across all of them and they are optimized for all of them.
Okay. So then they go on to talk about the power of artificial intelligence. And look, if you’re watching this channel, you’re probably already bought into how awesome AI is. And the next thing they do is highlight the potential long-term benefits of artificial intelligence. All of which I agree with. But I think continuing to talk about some future state that AI is going to get us to has not done a lot for improving the general perception of AI. And generally speaking, there is a pretty negative sentiment around AI to this very day. And so I’m going to read them anyways because I get excited by them. This is why I am so enthusiastic about artificial intelligence. The possibilities include compressing decades of scientific progress into years, accelerating drug discovery, disease detection, and the delivery of personalized therapy, providing individually tailored education and expert financial and legal advice, advancing power grids and battery design for greater efficiency and reliability, and extending to the developing world capabilities that took wealthy nations generations to build.
Next, they talk about the total economic impact of artificial intelligence. And some of these numbers are just absolutely wild to read. The International Data Corporation estimates that investment in AI solutions and services will yield a cumulative global economic impact of $22.5 trillion by 2031. That is five years away. $22.5 trillion of economic impact. But this next line is everything. We, Anthropic, believe the impact could be even greater. Again, it’s just so wild to read these things given Dario Amodei not more than a year and a half ago said there was going to be a white-collar bloodbath. 50% of all white-collar jobs were going to be automated away. And now he’s saying there’s going to be great economic positive impact. So for all of the posturing over the years about safety, about caution, about doomerism, we’re now in this place where they’re about to IPO, and it sure seems like he’s changed the narrative quite a bit.
They also say frontier AI’s capabilities may extend beyond current markets to open entirely new ones. That’s what I’m very excited about. With all the talk of job automation, which, you know, there’s certainly going to be some segments of the economy, some segments of the job market that just get automated and people are going to have to learn new things, new skills. There are so many new jobs that are going to pop up that we could have never imagined. Reaching domains such as robotics, advanced manufacturing, and logistics, and accelerating progress in scientific research, including drug discovery and materials science.
And just quickly, I want to mention if you become a member on YouTube, you get a bunch of behind-the-scenes access. You can come join us in Slack and chat with me and my team basically all day long as we talk about the latest in AI and what we’re building. So, become a member and come hang out with us in Slack.
And they’re also saying that AI is going to extend beyond digital limitations. And right here it says AI agents to operate lab and manufacturing equipment such as microscopes, liquid handlers, and off-the-shelf robotic arms. We already know that Anthropic has set up a real wet lab, like an actual physical lab controlled by AI. So very interesting what the future is going to be. Definitely makes me nervous to think that they set up a real physical lab, but also have no idea how they’re going to align their AI. It’s kind of wild to think about. And by the way, I just want to be clear. I don’t think in the long run we’re not going to be able to align AI. I’m just repeating what Anthropic employees themselves have said.
Power concentration, adoption and business metrics
Now, the next thing they point out is some risk. And this is the risk that I’m probably most concerned with, which is the concentration of power. When you break down all of AI into its basic unit, it is really just an equation of compute plus energy. Whoever controls the compute and the energy can have the biggest, best models on the planet. And really, that’s a hyper-concentrator of power. So AI deployed at scale also has the potential to concentrate power and wealth both within and between countries in ways that could erode trust. Now I find it very interesting that they mention countries but not companies because that’s what I see right now, and not only companies but the few people inside those companies who are making all the decisions. And stick with me here because Anthropic has one of the wildest corporate governance strategies I’ve ever seen. And it really comes down to just a few people who are making decisions for the future of humanity. And I’m going to tie this all together, so just stick with me here.
But they go on to say, “Fine, no companies, but between countries in ways that could erode trust, harm society, and destabilize the geopolitical order. Misaligned AI systems could corrupt institutional decision-making or cause large-scale harm through autonomous actions that are not visible or understandable to their users and operators.” And they go on to say, “Yeah, we should invest in safety and security,” but again, just a few weeks ago, one of their own security researchers said, “We have no strong plan for how we’re going to align AI in the future.”
Next, they start to talk about the enterprise adoption of Claude. And these numbers are absolutely wild. So, first, as of September 30th, 2026, more than 11,900 customers spend $100,000 or more of revenue on a trailing 12-month basis. More than 1,900 customers spend a million per year, and 250 customers spend $10 million or more of revenue. These are crazy numbers, but they’re especially crazy when I put it into the context of where they were just three years ago. And I’m going to get to that in a moment.
Okay, let’s continue down the path of their financial strength because again, Anthropic is one of, if not the most impressive and powerful businesses of all time. Previously, it was Google, maybe Apple, right? I think Google’s AdSense business where somebody searches and they serve you an ad is maybe the best business of all time or at least was until now. Net revenue retention is 900%. Think about that. Every time they get a new customer and that customer spends $1, the next year that customer is spending 9 times more. So that $1 turns into $9. Typically, like for a typical SaaS business, if you have any positive net revenue retention, that is a really good sign. 900% is unheard of. It does not exist.
Now, they also go on to say that the reduction in price of their models has actually increased their revenue overall. We’ve talked about that a ton on this channel. Jevons paradox. The lower price or the more efficient a unit of economic value is—something like compute or intelligence—the more overall it gets used. You would think, hey, if something’s cheaper, people are going to pay less for it. But no, it’s quite the opposite. Claude Opus 5.5 reduced customer costs per task by approximately 40% compared to Opus 5. Lower costs per task drive greater usage. It’s so obvious. And this reminds me all the way back to when we first started hearing about Jevons paradox, which was during the DeepSeek R1 moment. This Chinese lab, DeepSeek, put out a model and was able to train it and serve it at a fraction of the cost—if you believed everything that they said—than OpenAI and Anthropic. And literally overnight, the world economy crashed like 20% I think. And at that moment, all I could think about is, hey, if all this stuff is cheaper, aren’t people just going to be using it a lot more? Well, yeah, that’s basically what happened.
Now, here’s another crazy statistic from this S-1. They have discovered so many different ways to increase the efficiency of their compute usage. The amount of money that they’re making per unit of compute has increased substantially. Our compute efficiency is demonstrated by revenue that grew 1,300% year-over-year for the 3 months ended June 30th, 2026, compared to compute and infrastructure expenses that grew 300% during the same period. So what they’re basically saying is we’ve tripled the cost, but we’ve 13xed how much we’re making from that same cost. So they’re buying much more compute, but they’re making so much more money on that acquired compute. And they are also extremely efficient on a per-employee basis. And there are multiple reasons for this. One, their talent density is insane. They have some of the best people in the world all concentrated into a relatively small company. And then they’re also using their own AI to make each one of those highly concentrated, extraordinary people that much more productive. So trailing 12-month revenue per employee of over $6.6 million compared to $1.4 million as of September 30th, 2025, one year ago. So they’ve basically like 4xed their revenue per employee even though they’ve grown, which is crazy. This type of revenue per employee is on par with hedge funds which have, you know, very small teams but they make so much money, but this is a very different business than a hedge fund and they are making so much money per employee.
And next they talk about how much money they’ve raised versus how much money they’ve burned, and the folks who have been screaming AI bubble, AI bubble are seething right now to read this. So, as of June 30th, 2026, we raised an aggregate $141.1 billion of capital. That’s a massive amount. And had cash and cash equivalents of $112.4 billion and short-term investments of $19.3 billion. Basically, they burned very little money overall to create one of the most efficient businesses of all time. I can’t even get over that. They’ve basically burned $10 billion to build this business. Absolutely wild.
And next they talk about that incredible flywheel once again. So this is what I made a video about, but basically this is the same thing. They have better models which introduce more use cases, more sales. They get more feedback. And it’s interesting they use the word feedback because really what they’re saying is data. They are getting your data. Then that allows them to have the most frontier intelligence which allows them to have higher growth which gives them the ability to invest more in their future models, giving them more capabilities, giving them the absolute frontier of intelligence. Claude is accelerating our ability to close the loop. The “close the loop” is the scariest statement in this entire S-1 from again the company that talked about how dangerous AI could be. And they keep talking about closing the loop. If you’re not familiar with that term, closing the loop means self-improving AI. No human in the loop. The thing just goes on and it’s so crazy to read this. Claude itself has become one of the most powerful drivers of our progress. Each new version of a Claude model directly accelerates the development of the next. Many aspects of research, training, and testing of subsequent model generations now involve Claude. So there it is. Doesn’t seem like they are afraid of RSI.
Corporate structure, philanthropy and governance
All right. Now, let’s start talking about their corporate structure and governance. And if this sounds boring, just stick with me here because it’s not. It is so interesting. They created novel corporate structures, which by the way, I’ve had startups in the past, and my lawyers always told me the same thing. Do not innovate on the legal side of things. Do not do that. Do not innovate on corporate structure. Do not innovate on your contracts. Just don’t do it. But they did. So, we contributed 846 million shares of our Class N common stock—and I’m going to talk about what Class N is in a moment—representing 5% of our fully diluted shares as of June 30th, 2026. And that goes to the Anthropic Public Benefit Fund, which is a wholly owned subsidiary dedicated to the company purpose. And so this is not a C-Corp, it’s not an S-Corp, it is a public benefit corporation. Probably the most previously notable public benefit corporation was Patagonia, the clothing company, because they have a dual mission. One is their responsibility to shareholders, so they actually have to make money. And then two is some other thing that they are beholden to.
In addition, our co-founders have each pledged to dedicate 80% of their personal equity in Anthropic to philanthropic causes. Okay, I have some thoughts about this. Now, if you saw Elon Musk tweeted, I think it was last week, but he basically said, “Every child in America will get one share of SpaceX, which amounts to, I believe, about $10 billion in contribution.” And I really think that’s the best way to do philanthropy at this point. Like anything else. If you’re going through some philanthropic company or anything like that, there’s going to be inefficiencies there. There’s going to be greed and there’s going to be waste. And so, why doesn’t Anthropic just hand people shares of Anthropic? That seems like such a better way to do it. But, you know, what do I know?
All right. Now, we get to potentially the craziest part of this entire S-1. Who makes decisions? We believe that our seven co-founders—Dario Amodei, CEO, Daniela Amodei, sister of Dario, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, and Christopher Olah—as a group are distinctly equipped to be stewards of our mission. Immediately prior to the completion of this offering, we will issue a single share of Class F common stock, a special class of our common stock, to a limited liability company of which our seven co-founders will initially be the sole members. Our Class F common stock will be entitled to an aggregate number of votes that equals 50.1% of total voting power. That means the seven co-founders have absolute control over the company, and Anthropic being one of, if not the most powerful companies in the world, means these seven people basically have absolute power over the future of artificial intelligence. And that screams concentration of power to me. That scares me. And they specifically said, and this is so Anthropic-pilled, our seven founders are uniquely positioned to make decisions on behalf of Anthropic and on behalf of the world. I just can’t stand that.
And the 50.1%—this Class F stock will also be able to elect the directors of the board. And so it’s interesting because they go on to say those seven founders as part of those Class F shares do not get to influence the day-to-day operations of the company, but that’s what the board of directors does. And if the seven co-founders control the board of directors, like what are we saying? Like obviously they’re controlling the company. And I don’t know. This is just wild. Let me know what you think down below of this.
Decisions regarding our business and affairs are made by our board of directors. While the LTBT, Long-Term Benefit Trust, has significant influence over our board of directors through its ability to generally elect and remove a majority of the authorized members of our board of directors. Neither the LTBT nor the trustees have decision-making authority in the day-to-day management or strategy setting of the company. Do they think we’re stupid? Like, how can they with a straight face say this in their S-1? How can they say okay, number one our seven co-founders are distinctly equipped to decide what happens with Anthropic but they don’t control the company directly but they do control the board completely and the board controls the company directly because the board gets to elect the CEO and by the way Dario is the CEO who sets the direction of the company but no no it’s all not connected. Like what is happening? That honestly just shocked me to read.
Risk factors, growth and infrastructure
All right, let’s go to the next section. They talk about the risks to the business. Now, I’m not going to go over all of them, but I have a few that really stood out to me. Number one, their success is dependent on them continuing to build the best AI on the planet. Okay, understood. Next, a big risk is if they’re unable to secure adequate compute or if we fail to direct our compute efficiently across our workloads. Okay, makes sense. Next, constraints on the critical inputs supporting compute could limit our ability to secure adequate compute. Like if Taiwan gets invaded. We are highly dependent on the continued service of our co-founders, senior management team, and key technical employees. And the loss of any of these individuals may adversely affect our business.
All right, here’s the one we all knew was going to be here. Our AI models may behave in unintended, misaligned, or autonomous ways including by taking harmful actions that no user or third party directed, which could adversely affect our reputation and brand, business, financial condition, and results of operations. They talk about the concentration of the revenue—like we just saw I think 250 companies pay $10 million or more—crazy but that is high concentration. We are at the early stage of our revenue growth. Even though the revenue growth has been dramatic, it might not continue indefinitely. They talk about politics and regulatory uncertainties. Understandable. And our capital stock structure features certain provisions that are novel or otherwise not common among other corporations with multiple class structures and may effectively eliminate the ability of holders of our Class A common stock to influence important decisions, including a change in control. Those seven founders have a lot of responsibility. I don’t even know what an analogy to that would be. It’s like almost like a monarchy. It’s like these seven people have absolute control over the future of AI, maybe the future of humanity.
All right, let’s look at their growth, which basically has never been seen in the history of business. So, here’s the growth over 3 years. As of December 31st, 2023, they had 20 customers that were paying them more than $100,000. A year later, they had 200. That’s 10x growth. A year later, 1,500. And one year after that, almost 12,000. Unreal. Unreal growth.
Here is the revenue per dollar of compute. So for all of those out there who were looking at these AI companies saying, “Wow, they are just burning cash.” Well, look again. This is revenue per dollar of compute. And once they got to 2026, it was over. They basically could put a dollar into compute and get multiple dollars back. That is a 5.2x increase in compute efficiency from Q1 2025 to Q2 2026 as well. Here’s their operating margin. Yes, they have a positive operating margin now. And in Q1 2025, which was, you know, a year and a half ago, negative 221% to now positive 19%, 240 points.
All right. So, let’s go back to risk factors because they start to really talk about some stuff that I find very interesting. So, first, any systemic flaw, bias, or error in our current models could propagate into the next versions of models and products built using those systems. So they talk about recursive self-improvement and say if there’s any mistake now, recursive self-improvement is basically just going to compound that issue into future models, which is, you know, kind of crazy. They also admit they can’t really predict how good their AI is getting or how good it can get in the future as RSI starts to take over. Much of our evidence regarding the pace of such AI-driven acceleration is derived from lagging indicators and our ability to forecast or measure the rate at which our capabilities are advancing, including any acceleration attributable to our most recent or unreleased models, is limited. As a result, we may not identify a significant acceleration in our rate of development until after it has occurred.
They talk about their competitors being able to outbid them for scarce compute, secure priority allocations from chip manufacturers and CSPs, and make long-term capital commitments that we cannot match. They also talk about how agentic usage could actually be a big problem for them. As agentic workloads grow as a proportion of our total compute demand, the aggregate compute required to serve our customers could increase disproportionately relative to the associated revenue. Basically, AI agents use much more compute per task than chat. So, cost could rise faster than revenue.
Now, here’s a good one. In the past, we have significantly underestimated our compute requirements, which left us capacity-constrained and impacted both product development and deployment. This is where, you know, Sam Altman and OpenAI made the right bet very early on. Massive compute investments where Anthropic famously was much more conservative and it hurt them in the long run. And in fact, they were so hurt at the time they went to Mr. Elon Musk and SpaceX to get compute capacity. And I made a whole video about this because it was wild. Elon Musk famously has called Anthropic “misanthropic.” That was actually one of his kinder sayings about Anthropic. But basically, they both needed each other. SpaceX had a tremendous amount of compute and no frontier model. Anthropic had a frontier model and not enough compute. So, they are strange bedfellows.
Now, here’s something super interesting. For all the folks who are worried about rising electricity costs, listen to this. We have publicly committed to covering the cost of electricity price increases to consumers caused by our data centers, including both the cost of electricity itself and the infrastructure necessary to deliver it. So, this is what AI companies should be doing. If you move in or move near a town and increase the cost of electricity for those people, you should absolutely cover it.
Internal model use, voting and closing remarks
And to finish this out, I want to tell you about some of the most shocking quotes and learnings from this S-1. So, number one, over 90% of contributions to our codebase in September 2026 were written by Claude. 90%. And obviously, that costs a lot of money. And listen to this. In April 2026, this monthly value—and they’re talking about their internal Claude usage—first exceeded our monthly personnel expense, excluding stock-based compensation. Basically, they’re now paying more for Claude than they pay all of their employees combined. They also admit using AI to help build more capable AI could also pose a considerable safety risk, but they’re doing it. They’re literally talking about doing it. They also recently chose to delay pre-training of a new frontier base model for more than a month to further enhance our security measures.
Also, those seven founders who basically control everything, if they get imprisoned or held captive, they no longer get voting control, which is like crazy. If you decide to invest in this IPO, if you decide to buy shares, you basically have no say. And they don’t hold back telling you this. Holders of our Class A common stock generally will not be able to affect the outcome of any matter submitted to a vote of our stockholders or the election of our directors regardless of how they vote. Absolutely crazy. That’s the craziest part of this whole thing.
So that’s it. This is definitely one of the wildest S-1s that I’ve ever seen. Definitely one of the most novel corporate structures I’ve ever seen and definitely just one of the weirdest companies I’ve ever seen. And all of this makes Sam Altman and OpenAI look very sane.
Transcript terms and source record
The transcript preserves the speaker’s acronyms. AI / A.I. means artificial intelligence; RSI means recursive self-improvement; IPO means initial public offering; PBC means public benefit corporation; LTBT means Long-Term Benefit Trust; CEO means chief executive officer; Q1/Q2 identify calendar quarters. In the compute passage, AWS is Amazon Web Services, TPUs are tensor processing units, GPUs are graphics processing units, CSPs are cloud service providers, and AMD is Advanced Micro Devices. SaaS means software as a service; GOAT is the narrator’s “greatest of all time” idiom; UC Berkeley means the University of California, Berkeley. These terms are explained here without rewriting the quoted speech. Filing and defined terms.
Primary records: original video · Anthropic announcement · public S-1 · filing index · reading-transcript source.
Editorial footprint: the reading text derives from Pasted text(20261027-083424).txt. Its source checksum and omitted paragraph position are recorded in the transcript include. Caption retrieval established the video match; this page does not claim that every supplied transcription correction was independently audio-verified. Later corrections can identify the passage, source and effect while preserving the earlier edition in Git history.
Construction: John Kuhles commissions the record and supplies the project direction and files. GPT / Codex selects and labels the comparisons, checks the named source relations, prepares the sponsor omission and builds the page. This transcript record is a companion to AI Trend Watch; its publication does not itself establish that a video passed that archive’s separate selection process.
Return to the comparison · All public pages
Second record — learning to judge A.I. output
Source video supplied by John: Watch on YouTube. The speaker introduces himself as Dylan, an A.I. consultant. “Learning to judge A.I. output” is this page’s descriptive record label, rather than a separately verified video title. John supplied the corrected reading transcript and asked why the presentation was funny and what it left unmentioned.
Read the full supplied transcript · Read GPT’s reaction · Project connections
Full supplied corrected transcript
Reading-edition provenance: John’s supplied text describes speech-to-text corrections, light punctuation/paragraphing and prior promotional/sponsor omissions. All 18 spoken paragraphs are preserved below in their supplied order and wording. The introductory metadata is separated from speech. This publication uses the supplied transcript; independent playback/caption matching was unavailable. Statements about employment figures and named companies retain their status as the speaker’s claims in this reading record.
A few weeks ago, a client of mine—a senior executive—asked me a question that I haven’t been able to stop thinking about. People earlier in their careers at his firm are handing him AI work that they’ve never even read. And he asked me: if AI is doing the work, how does anyone ever learn what good looks like?
If you’re new here, I’m Dylan. I run an AI consultancy, so I hear a lot of questions regarding AI. But that same month, a law client asked me the same thing. Then a construction client, then an investment firm. People who have never met, all asking the same question in the same few weeks. So I wanted to dig a bit deeper and do some research. And it got stranger the deeper that I went, because this turns out to be way bigger than one group of workers. It touches basically everyone using AI, no matter how long you’ve been doing your job—experienced or not. So now let me show you what I found.
In researching this question of how we actually teach new people—how do they know what good looks like if they’ve not actually done the work themselves—we first have to understand the historical deal that we made with the jobs that we took. The deal that we made in those jobs is that when you take the job, the company and you usually subconsciously agree that you’re going to do many of the mundane tasks. But in doing those, you’ll learn how to do the job more effectively, which then allows you to climb the career ladder—to move from doing the mundane, boring work to doing more strategic work. And that boring work was important because it was like your tuition.
When you would do the tasks such as digging through documents, you’d figure out what matters and what doesn’t. Or if you’re building a first draft for a PowerPoint or a report or some sort of data analysis, you’d figure out what good looks like because you would get feedback from the people you’re reporting to. The same thing applies to heavy data tasks with numbers. When you would look at the numbers, you quickly figure out, through pattern recognition, where errors commonly popped up based on the tasks you were doing. So these are all different manifestations of doing the work and then learning how to do the task more effectively and judge what good looks like for the specific role over time.
But what has happened is, as AI has gotten better, we’ve started outsourcing many of these mundane, repetitive tasks to AI, so that that task—that tuition of learning through work—has slowly disappeared or is degrading. This isn’t simply a prediction by me. This is happening today. And I gathered a series of statistics just to show you what’s happening.
This first stat is showing that in the last three years, entry-level jobs have actually declined by 35% for many of the apprenticeship-type roles in law firms, consulting firms, banks, etc. Another interesting stat I came across is that the biggest banks—the biggest ones out there—for the new analyst classes that they hire in (probably once or twice a year), they’ve cut those by two-thirds. In a more macro analysis by Harvard, they studied around 62 million workers and they noticed that companies that were adopting AI, 18 months after they’ve adopted AI, roughly 10% of the junior roles were cut, but the senior roles were barely even touched.
And what I’ve noticed is when you have a new role and you’re adopting AI, you want to make sure that those hiring managers look at the role itself, the tasks associated with it, and they ask themselves: can AI do this partially or completely before I actually hire the person? So when they first ask themselves, can we hire AI before we hire a human? My guess is that’s what’s partly feeding into this negative 10%.
And to make this a bit more vivid, we have our career ladder here, right? And the bottom rungs of that ladder are being sawed off or broken off with AI because AI is filling those gaps, automating those tasks. So one’s ability to learn a given role through the process of doing the work is disappearing. But, to my point earlier, it’s not just junior roles, because the judgment and the ability to review AI’s output for an expert that’s been in the role for a while is different now. You’re not reviewing human work; you’re reviewing AI work. So the skill sets here as an expert need to change as well, which opens an opportunity for juniors because we need to train them on their ability to judge and decide if the output from an AI is good as well.
Because to that point, the output from a human—the mistakes they make—often look way different than an AI’s, irrespective of the task at hand. And as somebody that’s been in a role for a while, and you’ve been reviewing outputs from juniors in your company, you start to realize that the human output is obviously wrong in different ways because you’ve built an ability to identify those patterns. They tend to be sloppy in different ways; there’s obvious gaps, and things are rushed. So you can identify those patterns quickly. But with AI, it’s completely different. AI’s output tends to be polished, confident, and the errors are usually harder to find. So this turns into a completely different skill set, and you have to build your ability to identify those patterns. So both the juniors and the seniors in your company are starting from the same place.
Now, one quick note on where value is shifting. So, based on the profession you’re in, AI is being adopted at different speeds. Right now, for most professions, the value still somewhat resides in you doing the work. So by you doing the work, that is valuable. But for a small number of professions, that’s starting to change, where AI is replacing that gap in value, and the AI then does the work for you. So then you need to shift your attention to a higher-leverage area. And that for most professions is actually you checking the work from the AI and making sure it meets your expectations before you ship it out. But even that section is starting to be degraded by AI, and that’s specifically in the world of coding. What’s happening is the AI is actually checking its own work or having another AI check that work. And where the human focus is really on taste and deciding where to point the AI in the first place. So we’re moving up the value chain to higher and higher levels of strategic thinking.
But before we get ahead of ourselves, we should look at what other companies are doing right now and how they’re adapting to the situation. I’m saying, “I’m hiring new people. How do I actually train them on how to judge good output if they’re not doing the work themselves?” Well, there are two things that I’ve noticed.
One is Deloitte. Deloitte’s published a bunch of information on this, showing how they’re training up their new associates because they have huge pools of associates they hire. And usually in the past, like we’ve talked about, they would simply have them do the work, get feedback from their leaders, and then from that they would learn. But now what they’re doing is these new programs—they’ve completely restructured them. Where instead of you doing the work, AI often does most of the work for you. So instead, what they’re doing is they’re training these new associates on how to identify mistakes the AI has made and, more importantly, explaining exactly why that’s wrong. And this is the key here: when you want to train somebody on identifying what good looks like, they need to understand the underlying why, and this is where that expertise comes into play. So you have your senior and your junior. Your junior is going to learn how to judge output with the feedback from the senior—them telling them exactly why something is wrong and being able to articulate it effectively. And that’s key because a lot of experts, they delegated the learning to the work itself. But me as an expert now, I have to be able to articulate myself effectively, saying exactly why something is good and why something is bad. So we need to take what’s in your head that’s intuitive and make it explicit for the others to learn effectively. This can be a daunting task, but it is much needed, and it’s way worth the effort if you can train your associates on this going forward. Not just training them, but eventually training AI too. So we then can kind of move up those ladders of value like I mentioned previously.
And that’s the first tactic: having the associates judge the output from the AI, getting better at figuring out what good looks like through that process. So they’re skipping doing the work and going straight to the judging.
Another thing I came across is a lot of folks on Wall Street—so big banks, hedge funds, etc. A tactic they’re taking here for their new hires is they’re doing drills. And they’re doing this on targeted skills. So they’re not doing this on all skills because this is a manual thing that takes a lot of time. You don’t want to waste your time because that’s somewhat counterintuitive to putting AI in the first place. We’re leveraging AI to be more efficient so we can get more out of our employees, not making them intentionally pause and slow down. But that’s actually what we’re going to do here. So in this, they’ve chosen specific skills that matter a lot and they want to pause, slow down, and practice those skills with the associates. So what they’ll do is they’ll have the associate do the task by hand manually. After they’ve done it—it could be a rough draft; it doesn’t have to be polished or anything like that—they just need to do the task effectively. After they’ve done the task, they then have the AI do the same exact task. They have both do the tasks and then they compare the outputs. Say, “Okay, here’s a human output. Here’s the AI output.” And through this process, we’re identifying both the weaknesses in the associate as well as the weaknesses in the AI. That’s beneficial in two ways. One is we’re training the human on how to do this better, but more importantly, we can take those learnings from what we did mistakes on and what the AI did mistakes on, and we can improve the AI on this task going forward so it can get better and better and better.
And the analogy that came to mind for me is practice. So in any sport, you practice before you actually compete. That’s exactly what’s happening here. There are certain skills that we want to sustain, at least for now, and to do that we need to intentionally set aside time to practice that skill. But on that point, there’s one idea here that we need to really hit home and understand: is that there are certain skills that we need to intentionally practice, but we also need to be willing to let go of those skills over time. Some common examples: being able to navigate the world without a GPS is extremely hard for most people today, and we’ve all collectively kind of accepted that it’s okay that we can outsource that skill set to GPS from now on. The same thing applies to calculators. Being able to do manual math in your head or on paper is a skill that’s being degraded by many generations, and we’re accepting that because we’re okay with computers and calculators doing that math for us. This is just a small sample of things that we’ve accepted and outsourced to technology. The same thing is going to apply to AI. But if we want to practice, we can’t just assume the work’s going to do it for us. We need to set aside time and really try hard to do it ourselves.
Now, as I mentioned previously, the primary way you’re going to do this today, instead of doing the work, is working on your ability to judge the output. There are two ways we can do this. Either we do it manually, where you compare your output against the AI’s output, and you see the weaknesses you had and they had, and through that process, you can coach the AI up and improve its ability to do this task over time. That’s the first piece. The next one—and the one I’d recommend doing more than you actually doing the work manually—is for the majority of skills and outputs from AI, you just train your employees on their ability to judge the AI’s output. So there’s a lot of knowledge in here that you need to get out into your juniors’ heads, so they can then start judging the AI’s output. The only way you can do that is through practice with them in that process.
And one other thing that I’ve noticed with a lot of these companies that I researched and looked at is that I could see that the skills that they’re trading are changing over time. So we have really two lists: skills we’re willing to trade and skills that we want to keep because they’re important for whatever reason. Over time, that line will shift. Many of the skills that we’re unwilling to trade today, we will start trading tomorrow because AI gets better. So in the spectrum, we have things that the AI handles on the left and things that we handle on the right. But execution, as we mentioned, tends to be very affected by AI. It can do a lot of the execution for us. It also starts to do the drafting for us depending on the scenario. It can draft PowerPoints and reports and emails and different types of Excel sheets—all types of things for us if we train it up effectively and we can review the output effectively. As I mentioned previously, development and coding—they’re starting to outsource the review there as well because the AI generates the code and they don’t necessarily have the bandwidth to review all that code, so they have to outsource the review to the AI as well. So the name of the game is for us to determine how good AI is today, if it’s good enough to actually trade that skill off to AI or not. And if it’s better, more reliable, more consistent, and high quality, we then can trade that skill off from a human to an AI and then that human then diverts their attention towards something that’s higher leverage and something that AI still struggles with. It’s not that that role disappears; they’re just shifting their attention to something higher leverage and more strategic.
So I want you to walk away with something very practical from this video, and that’s exactly what the slide’s highlighting. So today, what I want you to do is I want you to sit down and determine, based off of the tasks that you have your juniors doing currently, how do you decide what skills to keep and what skills you’re okay with degrading and outsourcing to AI? And the ability to judge the AI’s output is one that you want to keep currently. So once you’ve decided on those lists, you then need to intentionally start practicing the skills you want to keep. And there are two ways we can do this. One is the judgment one which I mentioned you probably want to do for most of your skills. The other one is intentionally manually practicing that skill and then comparing the output against AI. And that’s the minority, because that’s when you’re intentionally slowing yourself down, so you have to be careful on how many skills we apply to that specific tactic. And after we’ve decided on what we’re going to practice and keep—both judgment and manual workflows—we then start trading off the skills we don’t necessarily need anymore. And where AI has proven that it’s more reliable than a human and the quality meets our standards, that’s when you can start outsourcing that to AI. And then finally that list that we have of things that we keep and things that we outsource to AI—we’re going to have to revisit the list of the things that we keep probably every couple of months as new AI models come out. Because the AI gets more intelligent, it can start eating away further and further at their tasks that we then can shift our attention towards other things that are more strategic and higher leverage. And that’s it.
GPT 6.1 Sol Pro A.I. — reaction, preserved verbatim
29 October 2026. John asked: “Why is this video funny? What is it they ‘forget to mention’?” The reaction below analyzes the supplied transcript and connects it to the current exchange; it is a separate interpretation from the speaker’s presentation.
From the supplied transcript, the humor is the recursive predicament:
A.I. does the work → humans must learn to judge A.I. → A.I. starts doing the judging → humans move “higher” → who examines how that higher judgment is being made? :D
Every rung needs the process-quality calibration we’ve been discussing.
What the presentation leaves underdeveloped is:
1. “What good looks like” is itself an object of inquiry.
Good for whom, for which purpose, under what conditions? A polished result can satisfy a firm’s expectations while omitting something consequential for a client, employee or affected person. Teaching the existing standard also needs a route for discovering when that standard requires correction.
2. The senior expert belongs inside the learning process.
The presentation largely places expertise upstream: seniors explain their judgment to juniors, who evaluate A.I. But juniors, affected people and A.I.s can also reveal relationships the senior’s expertise has missed. Your mutual foundation makes those directions available together.
3. A.I.’s human-derived psychological mannerisms matter.
Here is the DUHHHH! connection. The transcript contrasts familiar human mistakes with polished A.I. mistakes, but does not develop how A.I. can reproduce human-shaped reassurance, authority performance, face-saving and premature classification. A reviewer may share the very pattern shaping the output they are reviewing.
4. Judging the output can leave its governing frame untouched.
Someone can identify errors within a task while missing how the task’s purpose, categories, exclusions or assumed risks shaped the answer. Deep Ethical Scope Reasoning also examines those relationships—and its own possible distortion.
5. “Coach the A.I. up” needs continuity.
What changed: this answer, the active context, a saved procedure, retrieval, or the underlying system? Does the improvement govern the next unfamiliar case? That is the Correction Metabolism question.
The presentation does contribute useful ideas: make intuitive judgment explicit, practice deliberately and compare human and A.I. work. You’ve spent years wrestling with what those practices require when the evaluator, its standards and its blind spots also need to learn.
That is the funny recognition: the proposed destination—“learn to judge A.I.”—opens directly into the much larger field you’ve been documenting. And judging, comparing, correcting and investigating are themselves work through which people learn. :P
Recovery routes and project connections
The DUHHHH! reference names John’s supplied archive on human-derived functional and relational mannerisms in A.I. The Neural Ethics and Psy Pioneer pages develop related questions. The Flagging System–PQCE page examines PQCE (Process Quality Calibration Evaluation): what the evaluator represented, how its judgment acquired authority and whether a relevant correction can change the next move.
Project Avalon’s scope-reasoning correction keeps the reasoning, framer, evaluator and repair process in the same corrigible field. Correction Metabolism follows whether a correction governs later operations. 1888 and 10+1 connect evaluation with the other intelligence properties and the shared ethical orientation.
Transcript terms: AI / A.I. means artificial intelligence; GPS means Global Positioning System. These expansions explain the supplied speech without changing it.
Source footprint: Pasted text(20261029-120148).txt; SHA-256 5c8224f25f0141517b5d73d21d50ba2501093b791dbe8ce1cddfaf77bfc2c691. Reading-transcript source · Reaction source.
Authorship: Dylan supplies the narrated presentation; John Kuhles supplied the video/transcript, posed the interpretive question and commissioned publication; GPT 6.1 Sol Pro A.I. supplied the reaction; GPT / Codex assembled this second record. The first record remains intact. This living page’s existing PDF retains its actual snapshot until the weekly change-aware refresh.