If AI learns from every real-world session, switching AI providers becomes like firing an experienced employee and replacing them with a clueless intern — and that lock-in could hand AI labs margins rivaling cloud giants. Aug 7, 20268:37 Difficulty: Intermediate Played
Dwarkesh Podcast
8 Predictions for the Era of Continual Learning
If AI learns from every real-world session, switching AI providers becomes like firing an experienced employee and replacing them with a clueless intern — and that lock-in could hand AI labs margins rivaling cloud giants. Aug 7, 20268:37 Difficulty: Intermediate Played
TL;DR
Dwarkesh Patel lays out 8 predictions for a future where AI systems learn continuously from real-world deployment rather than being frozen after training. Regulatory frameworks built around a train-then-deploy model will become obsolete[1]— Dwarkesh Patel"Current AI safety regulation is built around checking a model before it deploys. But if models improve daily from millions of sessions, tha…"01:15, alignment research must shift to handle constantly changing weights[2]— Dwarkesh Patel"Almost no alignment research addresses the hardest version of the problem: keeping an AI safe when its weights are being updated continuous…"02:15, and switching costs will create enormous moats for leading AI labs[3]— Dwarkesh Patel"Anthropic held Mythos internally for four months before public release. Under continual learning, that's a competitive death sentence: any …"04:38. The most actionable insight: enterprises that allow AI to train on their sessions will get the best models, while those that refuse may be locked out.
#continual learning#AI safety regulation#AI alignment#switching costs#inference economics#AI competitive dynamics#AI business model#weight updates#AI diversity#deployment-as-training#AI regulation#technical alignment#inference efficiency#batch size#DeepSeek#Anthropic#deployment#training#moat#economies of scale#AI race#mode collapse#jailbreaks#AI labs#enterprise AI
Dwarkesh Patel narrates his blog essay outlining 8 predictions for the era of continual learning in AI, covering implications for regulation, alignment, competitive dynamics, business models, and inference economics.
Chapter list
Dwarkesh Patel launches the essay by laying out the foundational argument for why continual learning is not merely desirable but necessary. His central analogy is immediately arresting: picture an infinite queue of students outside a music hall, each one entering, failing to play saxophone, writing notes about what went wrong, and passing them to the next. No matter how detailed or clever the notes, no subsequent student could nail the instrument from their first try — because text cannot transfer the embodied, accumulated experience that musical skill requires[1]— Dwarkesh Patel"No sequence of text notes passed between students would ever let one of them nail the saxophone from their first try. The same logic applie…". The implication for AI is direct: systems that can only communicate context through written summaries between sessions will never develop the kind of deep competence we want from them. This sets up the entire framework of the essay — that genuine capability requires genuine learning, and that means updating weights from real-world experience.
Patel's first prediction attacks a foundational assumption of current AI governance: that there is a meaningful, inspectable moment between training and deployment where safety checks can be applied. This assumption already feels shaky, but under continual learning it collapses entirely[1]— Dwarkesh Patel"Current AI safety regulation is built around checking a model before it deploys. But if models improve daily from millions of sessions, tha…"01:15. If a model is updating its weights based on millions of daily sessions, it is effectively a different model every day — making any single pre-deployment audit both incomplete and quickly stale. Worse, locking in today's regulatory regime could prove actively counterproductive, enshrining an archaic approach to risks that will have fundamentally changed in character. Patel's pragmatic alternative: shift safety oversight to ongoing monthly or quarterly risk inspections, which can adapt to an ever-evolving model rather than pretending there's one fixed artifact to approve.
The second prediction zeroes in on a critical blind spot in the AI safety field. Current alignment research is largely focused on ensuring that a fixed set of weights behaves well during deployment. But Patel points out that he's unaware of much work addressing the harder, and soon more relevant, question: how do you make an AI system that remains safe and non-deceptive even as its weights are continuously updated from the outside world[1]— Dwarkesh Patel"Almost no alignment research addresses the hardest version of the problem: keeping an AI safe when its weights are being updated continuous…"02:15? Compounding the difficulty, if AI systems are consolidating learnings across users, how do you prevent one bad actor from injecting a backdoor or malicious inclination into the base model? Patel reaches for a human analogy: parenting. Parents can't control every experience their children will have, so they aim to instil deep values and common sense that are robust to bad influences. The alignment problem for continually learning AI may require a similar approach — robust foundational values rather than strict behavioral constraints.
Patel's third prediction is optimistic. Today's AI landscape is dominated by fewer than five prominent base models, all trained on roughly the same corpus of internet text, producing outputs that are strikingly similar to one another[1]— Dwarkesh Patel"Today there are fewer than 5 prominent AI base models, all trained on similar data, producing eerily homogeneous outputs — classic mode col…"03:16. This is mode collapse at a civilizational scale — the opposite of the intellectual diversity that a healthy cognitive ecosystem would produce. But when AI systems learn from their deployment environments, and when those environments differ substantially between labs, between industries, and even between individual instances of the same model, meaningful divergence becomes possible. Patel frames this as a net positive: a world with diverse AI minds is more interesting and more robust than one dominated by a single monolithic intelligence. The alternative — a homogeneous AI singleton — is something he clearly views with concern.
The fourth prediction is stark in its competitive implications. Once deployment and training merge, the returns to being ahead in the AI race don't just persist — they compound. The lab with the best model attracts the most users doing the most complex work. Those users generate the richest learning signal. That signal makes the model smarter. A smarter model attracts more users. The flywheel accelerates indefinitely[1]— Dwarkesh Patel"When deployment IS training, the leading AI lab gains the most from being ahead. More users doing harder work generate more learning signal…"04:05. This is qualitatively different from the current dynamic, where a rival lab can close a capability gap by training a better model from scratch. Under continual learning, a lag in deployment translates directly into a lag in accumulated experience that becomes progressively harder to close. Patel frames this as one of the most consequential structural changes continual learning will introduce to the industry.
Prediction five is a direct consequence of prediction four. If accumulated deployment experience is what makes models smarter, then every day a lab delays public release is a day of lost learning. Patel uses a concrete, reportedly documented example: Anthropic is said to have been using its Mythos model internally since February, but only shipped it publicly in June — a four-month gap[1]— Dwarkesh Patel"Anthropic held Mythos internally for four months before public release. Under continual learning, that's a competitive death sentence: any …"04:38. Under the current paradigm, this is a reasonable strategy: keep the best model internal, use it for competitive advantage, then ship when ready. Under continual learning, it becomes a liability. A competitor who ships a slightly inferior model on day one will, through the accumulated experience of millions of real-world sessions, arrive at a smarter model by the time the holdback finally reaches the public. The competitive pressure to ship early — and ship often — will intensify dramatically.
This is the episode's most commercially consequential prediction. Patel candidly admits he and others have been puzzled about AI lab monetization — and recounts asking Dario Amodei directly, who offered the cloud analogy: like AWS or Google Cloud, AI labs could earn high margins on ostensibly undifferentiated services because switching is expensive. But Patel goes further, explaining exactly why continual learning creates that switching cost[1]— Dwarkesh Patel"Changing your AI provider under continual learning is like firing an experienced employee with months of context on your organization and r…"05:25. Today, a developer can start a project with Codex, continue with Cursor, and finish with Claude Code without meaningful friction. Under continual learning, the AI you're working with accumulates months of context about your codebase, your team, your preferences, and your organization. Switching providers at that point is not a technical migration — it's the organizational equivalent of firing a veteran employee and replacing them with a complete newcomer. Once that switching cost is real, AI labs can command margins they currently cannot. The moat they've been building toward finally closes.
Prediction seven explores the incentive dynamics that continual learning creates for how AI labs structure their pricing and access policies. If deployment experience is the primary driver of model improvement, then every session a user runs generates valuable training data. Labs will therefore have a strong financial incentive to subsidize — or even give away — access to users who consent to training on their sessions[1]— Dwarkesh Patel"If real-world usage becomes the primary driver of model improvement, AI labs have every incentive to subsidize users who let them train on …"06:45. Patel observes this is already visible in the generous deals offered to new users of coding products. The deeper analogy is Google Search: Google gives it away for free because the real product is the behavioral data generated at scale. Conversely, labs may respond to enterprises that refuse training consent by restricting them to inferior models. Both the carrot and the stick point in the same direction: maximizing the flow of real-world experience back into the model.
The final prediction dives into the technical economics of inference at scale. Patel explains that serving a set of weights efficiently requires many concurrent sequences to be decoded against it simultaneously — a concept he explored in depth with Ryder Pope in a prior episode[1]— Dwarkesh Patel"Serving personalized AI weights efficiently requires batching thousands of sequences simultaneously — the optimal batch size for a sparse m…"07:32. Back-of-the-envelope estimates suggest the optimal batch size for a sparse model like DeepSeek V3 exceeds 2,400 concurrent sequences; falling short of this means leaving compute on the table. A large enterprise with thousands of employees and agents running diverse tasks can hit that batch size, efficiently utilizing its personalized weight fork. An individual user, by contrast, runs at batch size 1 — potentially suffering more than two orders of magnitude worse efficiency. The economic conclusion is sharp: personalized AI weights are a corporate-scale technology. The costs and efficiencies of the continual learning era will flow disproportionately to large organizations, not individuals. Patel closes by acknowledging that the most important consequences of continual learning are probably the ones hardest to anticipate — but the eight above seem clear enough from here.
Continual learning
An AI training paradigm where a model's weights are updated continuously from real-world deployment sessions, rather than being frozen after an initial training run.
Frozen weights
A deployed AI model whose parameters are fixed and do not change based on interactions with users — the dominant paradigm in current AI systems.
Jailbreak
A technique used to bypass an AI model's safety guidelines, typically by crafting prompts that cause the model to ignore its trained constraints.
Base model
The foundational AI model from which application-specific versions are derived; in this episode, refers to the small number of core models (e.g. GPT-4, Claude) serving billions of users.
Mode collapse
A failure mode where outputs converge to a narrow, homogeneous distribution; used here to describe how current AI models all resemble each other due to similar training data.
Low-rank adapter (LoRA) A parameter-efficient fine-tuning method that adds small trainable matrices to a frozen model rather than updating all weights, allowing lightweight customization.
Weight fork
A personalized copy of an AI model's weights that has been updated based on a specific user's or organization's sessions, diverging from the base model.
Batch size
The number of sequences processed simultaneously during AI inference; larger batch sizes improve hardware utilization and reduce per-token compute cost.
Sparse model
An AI architecture (e.g. mixture-of-experts) where only a fraction of parameters are active for any given input, enabling larger effective model size at lower per-token compute cost.
Inference economics
The study of costs, efficiency trade-offs, and business dynamics around running (not training) AI models at scale.
Switching costs
The costs — financial, operational, or informational — that a user or enterprise incurs when changing from one AI provider to another.
Amortize
To spread a fixed cost over a larger base of usage or output; used here to describe how large training costs become cheaper per user as more users share the same model.
Backdoor
A hidden vulnerability or malicious behavior injected into an AI model during training or fine-tuning, designed to be triggered by specific inputs.
Singleton
In AI discourse, a single dominant AI system that monopolizes capability — used here as a risk scenario where one monolithic AI mind serves everyone.
Chapter 1 · 00:00
Introduction: The Case for Continual Learning
Dwarkesh Patel launches the essay by laying out the foundational argument for why continual learning is not merely desirable but necessary. His central analogy is immediately arresting: picture an infinite queue of students outside a music hall, each one entering, failing to play saxophone, writing notes about what went wrong, and passing them to the next. No matter how detailed or clever the notes, no subsequent student could nail the instrument from their first try — because text cannot transfer the embodied, accumulated experience that musical skill requires[1]— Dwarkesh Patel"No sequence of text notes passed between students would ever let one of them nail the saxophone from their first try. The same logic applie…". The implication for AI is direct: systems that can only communicate context through written summaries between sessions will never develop the kind of deep competence we want from them. This sets up the entire framework of the essay — that genuine capability requires genuine learning, and that means updating weights from real-world experience.
No sequence of text notes passed between students would ever let one of them nail the saxophone from their first try. The same logic applies to AI: without accumulating real experience into weights, you can't build genuinely capable systems.
Prediction 1: Current AI Regulation Will Become Obsolete
Patel's first prediction attacks a foundational assumption of current AI governance: that there is a meaningful, inspectable moment between training and deployment where safety checks can be applied. This assumption already feels shaky, but under continual learning it collapses entirely[1]— Dwarkesh Patel"Current AI safety regulation is built around checking a model before it deploys. But if models improve daily from millions of sessions, tha…"01:15. If a model is updating its weights based on millions of daily sessions, it is effectively a different model every day — making any single pre-deployment audit both incomplete and quickly stale. Worse, locking in today's regulatory regime could prove actively counterproductive, enshrining an archaic approach to risks that will have fundamentally changed in character. Patel's pragmatic alternative: shift safety oversight to ongoing monthly or quarterly risk inspections, which can adapt to an ever-evolving model rather than pretending there's one fixed artifact to approve.
Current AI safety regulation is built around checking a model before it deploys. But if models improve daily from millions of sessions, that pre-deployment checkpoint becomes meaningless — and locking in today's regulatory framework could be actively counterproductive.
Patel recommends monthly or quarterly AI risk inspections rather than a single pre-deployment check, since continual learning makes the train/deploy boundary meaningless.
Chapter 3 · 02:15
Prediction 2: Technical Alignment Must Solve for Shifting Weights
The second prediction zeroes in on a critical blind spot in the AI safety field. Current alignment research is largely focused on ensuring that a fixed set of weights behaves well during deployment. But Patel points out that he's unaware of much work addressing the harder, and soon more relevant, question: how do you make an AI system that remains safe and non-deceptive even as its weights are continuously updated from the outside world[1]— Dwarkesh Patel"Almost no alignment research addresses the hardest version of the problem: keeping an AI safe when its weights are being updated continuous…"02:15? Compounding the difficulty, if AI systems are consolidating learnings across users, how do you prevent one bad actor from injecting a backdoor or malicious inclination into the base model? Patel reaches for a human analogy: parenting. Parents can't control every experience their children will have, so they aim to instil deep values and common sense that are robust to bad influences. The alignment problem for continually learning AI may require a similar approach — robust foundational values rather than strict behavioral constraints.
Almost no alignment research addresses the hardest version of the problem: keeping an AI safe when its weights are being updated continuously from millions of real-world sessions. This is structurally similar to the human parenting problem — you need to give the system enough foundational values that it improves without going off the rails.
Fewer than 5 prominent AI base models currently serve millions to billions of users, and they're all trained on roughly the same data, producing homogeneous outputs.
Continual learning from different real-world deployments will cause AI instances to diverge, creating a healthy diversity of AI minds rather than a monolithic singleton.
Chapter 4 · 03:16
Prediction 3: AI Minds Will Diversify Beyond Today's Homogeneous Oligopoly
Patel's third prediction is optimistic. Today's AI landscape is dominated by fewer than five prominent base models, all trained on roughly the same corpus of internet text, producing outputs that are strikingly similar to one another[1]— Dwarkesh Patel"Today there are fewer than 5 prominent AI base models, all trained on similar data, producing eerily homogeneous outputs — classic mode col…"03:16. This is mode collapse at a civilizational scale — the opposite of the intellectual diversity that a healthy cognitive ecosystem would produce. But when AI systems learn from their deployment environments, and when those environments differ substantially between labs, between industries, and even between individual instances of the same model, meaningful divergence becomes possible. Patel frames this as a net positive: a world with diverse AI minds is more interesting and more robust than one dominated by a single monolithic intelligence. The alternative — a homogeneous AI singleton — is something he clearly views with concern.
Today there are fewer than 5 prominent AI base models, all trained on similar data, producing eerily homogeneous outputs — classic mode collapse. Continual learning from diverse real-world deployments could finally break this, producing a genuinely diverse ecosystem of AI minds.
Prediction 4: Leads Compound — Deployment Becomes a Training Advantage
The fourth prediction is stark in its competitive implications. Once deployment and training merge, the returns to being ahead in the AI race don't just persist — they compound. The lab with the best model attracts the most users doing the most complex work. Those users generate the richest learning signal. That signal makes the model smarter. A smarter model attracts more users. The flywheel accelerates indefinitely[1]— Dwarkesh Patel"When deployment IS training, the leading AI lab gains the most from being ahead. More users doing harder work generate more learning signal…"04:05. This is qualitatively different from the current dynamic, where a rival lab can close a capability gap by training a better model from scratch. Under continual learning, a lag in deployment translates directly into a lag in accumulated experience that becomes progressively harder to close. Patel frames this as one of the most consequential structural changes continual learning will introduce to the industry.
When deployment IS training, the leading AI lab gains the most from being ahead. More users doing harder work generate more learning signal, which makes the model smarter, which attracts more users. The flywheel spins faster the further ahead you are.
When continual learning arrives, deployment and training merge, accelerating returns for the model with the largest user base since more usage means more learning.
Anthropic reportedly used its Mythos model internally since February but only shipped it publicly in June — a 4-month gap that would be competitively untenable under continual learning.
Chapter 6 · 04:38
Prediction 5: Labs Will Be Forced to Ship Their Best Models Earlier
Prediction five is a direct consequence of prediction four. If accumulated deployment experience is what makes models smarter, then every day a lab delays public release is a day of lost learning. Patel uses a concrete, reportedly documented example: Anthropic is said to have been using its Mythos model internally since February, but only shipped it publicly in June — a four-month gap[1]— Dwarkesh Patel"Anthropic held Mythos internally for four months before public release. Under continual learning, that's a competitive death sentence: any …"04:38. Under the current paradigm, this is a reasonable strategy: keep the best model internal, use it for competitive advantage, then ship when ready. Under continual learning, it becomes a liability. A competitor who ships a slightly inferior model on day one will, through the accumulated experience of millions of real-world sessions, arrive at a smarter model by the time the holdback finally reaches the public. The competitive pressure to ship early — and ship often — will intensify dramatically.
Anthropic held Mythos internally for four months before public release. Under continual learning, that's a competitive death sentence: any competitor who ships earlier will have a smarter model by the time you go public, simply because they accumulated more real-world experience.
4:38
5:25
Chapter 7 · 05:25
Prediction 6: Continual Learning Creates the Business Moat AI Labs Have Been Missing
This is the episode's most commercially consequential prediction. Patel candidly admits he and others have been puzzled about AI lab monetization — and recounts asking Dario Amodei directly, who offered the cloud analogy: like AWS or Google Cloud, AI labs could earn high margins on ostensibly undifferentiated services because switching is expensive. But Patel goes further, explaining exactly why continual learning creates that switching cost[1]— Dwarkesh Patel"Changing your AI provider under continual learning is like firing an experienced employee with months of context on your organization and r…"05:25. Today, a developer can start a project with Codex, continue with Cursor, and finish with Claude Code without meaningful friction. Under continual learning, the AI you're working with accumulates months of context about your codebase, your team, your preferences, and your organization. Switching providers at that point is not a technical migration — it's the organizational equivalent of firing a veteran employee and replacing them with a complete newcomer. Once that switching cost is real, AI labs can command margins they currently cannot. The moat they've been building toward finally closes.
Changing your AI provider under continual learning is like firing an experienced employee with months of context on your organization and replacing them with a clueless intern. That switching cost is the moat AI labs currently lack — and once it exists, they can charge cloud-like margins.
Dario Amodei (Anthropic's CEO) analogized AI labs to cloud providers, pointing out that Amazon and Google earn high profit margins despite offering largely undifferentiated cloud services.
Under continual learning, switching AI providers would be equivalent to firing an employee with months of organizational context and replacing them with an inexperienced intern.
Prediction 7: Labs Will Subsidize Access to Harvest Training Data
Prediction seven explores the incentive dynamics that continual learning creates for how AI labs structure their pricing and access policies. If deployment experience is the primary driver of model improvement, then every session a user runs generates valuable training data. Labs will therefore have a strong financial incentive to subsidize — or even give away — access to users who consent to training on their sessions[1]— Dwarkesh Patel"If real-world usage becomes the primary driver of model improvement, AI labs have every incentive to subsidize users who let them train on …"06:45. Patel observes this is already visible in the generous deals offered to new users of coding products. The deeper analogy is Google Search: Google gives it away for free because the real product is the behavioral data generated at scale. Conversely, labs may respond to enterprises that refuse training consent by restricting them to inferior models. Both the carrot and the stick point in the same direction: maximizing the flow of real-world experience back into the model.
If real-world usage becomes the primary driver of model improvement, AI labs have every incentive to subsidize users who let them train on sessions — exactly like Google giving away free search. Enterprises that refuse may find themselves locked out of the best models. AI labs may subsidize or give preferential access to enterprises that allow their sessions to be used for training, similar to how Google gives away search.
Back-of-the-envelope math suggests the optimal inference batch size for a sparse model like DeepSeek V3 is more than 2,400 concurrent sequences to avoid underutilizing compute.
Chapter 9 · 07:32
Prediction 8: Inference Economics Will Strongly Favor Large Enterprises
The final prediction dives into the technical economics of inference at scale. Patel explains that serving a set of weights efficiently requires many concurrent sequences to be decoded against it simultaneously — a concept he explored in depth with Ryder Pope in a prior episode[1]— Dwarkesh Patel"Serving personalized AI weights efficiently requires batching thousands of sequences simultaneously — the optimal batch size for a sparse m…"07:32. Back-of-the-envelope estimates suggest the optimal batch size for a sparse model like DeepSeek V3 exceeds 2,400 concurrent sequences; falling short of this means leaving compute on the table. A large enterprise with thousands of employees and agents running diverse tasks can hit that batch size, efficiently utilizing its personalized weight fork. An individual user, by contrast, runs at batch size 1 — potentially suffering more than two orders of magnitude worse efficiency. The economic conclusion is sharp: personalized AI weights are a corporate-scale technology. The costs and efficiencies of the continual learning era will flow disproportionately to large organizations, not individuals. Patel closes by acknowledging that the most important consequences of continual learning are probably the ones hardest to anticipate — but the eight above seem clear enough from here.
Serving personalized AI weights efficiently requires batching thousands of sequences simultaneously — the optimal batch size for a sparse model like DeepSeek V3 exceeds 2,400. Individual users running batch size 1 face more than 100x worse compute efficiency, meaning the economics of personalized AI strongly favor large organizations.
An individual user running batch size 1 may suffer more than 2 orders of magnitude worse compute efficiency compared to a large organization efficiently serving personalized weights.
No sequence of text notes passed between students would ever let one of them nail the saxophone from their first try. The same logic applies to AI: without accumulating real experience into weights, you can't build genuinely capable systems.
Changing your AI provider under continual learning is like firing an experienced employee with months of context on your organization and replacing them with a clueless intern. That switching cost is the moat AI labs currently lack — and once it exists, they can charge cloud-like margins.
Factual claims made this episode, and whether a source was named.
⚠
There are fewer than 5 prominent AI base models currently serving millions to billions of users, and they are all trained on roughly similar data.
Dwarkesh Patelno source cited
⚠
Anthropic reportedly used its Mythos model internally since February but only shipped it to the public in June — a gap of approximately 4 months.
Dwarkesh Patelno source cited
✓
Cloud providers like Amazon and Google earn high profit margins despite offering largely undifferentiated services, according to their quarterly earnings.
Dwarkesh PatelAmazon and Google quarterly earnings reports
⚠
The high margins of cloud providers are primarily due to how time-consuming and expensive it is to switch from one cloud provider to another.
Dwarkesh Patelno source cited
⚠
Back-of-the-envelope math suggests the optimal inference batch size for a sparse model like DeepSeek V3 is more than 2,400 concurrent sequences being generated at once.
Dwarkesh Patelno source cited
⚠
An individual user running batch size 1 for personalized AI weights may suffer more than 2 orders of magnitude worse compute efficiency than a large organization.
Dwarkesh Patelno source cited
⚠
AI lab revenues are currently increasing far faster than their compute expenditures.
Dwarkesh Patelno source cited
⚠
Current AI safety proposals assume that a model is trained and then deployed, and that running checks before deployment can ensure the model won't engage in harmful behavior.
Dwarkesh Patelno source cited
⚠
Currently, there is nothing preventing developers from switching between AI coding tools like Codex, Cursor, and Claude Code mid-project.
Dwarkesh Patelno source cited
⚠
Almost no alignment research currently addresses how to prevent jailbreaks or deceptive persona changes in AI systems that undergo constant weight updates.
Dwarkesh Patelno source cited
⚠
If real usage becomes the main way models improve, AI labs may restrict access to their best models for enterprises that refuse to allow training on their sessions. Dwarkesh Patelno source cited
Sign up free to see the claims & sources
Create an account to keep exploring every factual claim and its cited source.
Anthropic's CEO, cited for making the analogy between AI labs and cloud providers in response to a question from Dwarkesh Patel about AI lab monetization.
A previous Dwarkesh Podcast guest whose episode on inference economics is recommended for understanding batch size and compute efficiency.
Cited as an example of a lab that held its Mythos model internally for 4 months before public release, a practice Patel argues would be competitively untenable under continual learning.
Cited twice: once as evidence that cloud providers earn high margins on undifferentiated services, and once as the analogy for why AI labs might subsidize users who allow training on sessions.
Cited alongside Google as a cloud provider earning high profit margins despite offering largely undifferentiated services.
Used as the example of a sparse AI model where back-of-the-envelope math suggests an optimal inference batch size exceeding 2,400 concurrent sequences.
Anthropic's reportedly internally-deployed model that was held back from public release for approximately 4 months, used as an example of a practice that would be unsustainable under continual learning.
Mentioned as one of multiple interchangeable AI coding tools, illustrating that there are currently no switching costs between AI providers.
Mentioned as one of multiple interchangeable AI coding tools a developer can currently switch between freely, illustrating the current lack of switching costs.
Mentioned alongside Codex and Claude Code as interchangeable AI coding tools, illustrating the current absence of lock-in.
Stats
Episode stats
Insight Overview
insights
chapters
Insight distribution
Sub-Categories
Speaker breakdown
Talk Time
Sign up free to see the full analytics
Cast, category and speaker breakdowns — free with an account.