
Sign up to save your podcasts
Or


Will AI take your job? The honest answer starts with what the research measures — and most of it measures tasks, not job titles. This episode works through the major studies, separates exposure from displacement, and lands on where the evidence is currently strongest.
On exposure: a landmark study from OpenAI researchers estimated that around 80% of US workers could have at least 10% of their tasks affected by large language models, with about 19% seeing at least half their tasks affected. The International Monetary Fund puts roughly 40% of jobs worldwide as exposed, rising to about 60% in advanced economies. Both measure exposure, not predicted job losses.
The episode surfaces a striking tension inside Gartner's own research. Gartner forecasts that by 2029, agentic AI could autonomously resolve as much as 80% of common customer-service issues. Then in April 2026, Gartner surveyed service and support leaders and found 85% were expanding what their human agents do rather than shrinking it — the same firm, the same industry, with forecast and current practice pointing opposite ways.
Worker-level research points the same direction. In a study of more than 5,000 customer-support agents, generative AI raised productivity about 14% overall and 34% for novice and lower-skilled workers, spreading techniques that had belonged to the most experienced staff. On the macro side, the World Economic Forum projects roughly 170 million jobs created and 92 million displaced by 2030 for a net gain near 78 million — though driven by multiple macrotrends, not AI alone — while 39% of workers' current skills change or go stale.
The sharpest current evidence concerns younger workers. Stanford researchers, using payroll data covering millions of US workers through June 2026, found employment for 22-to-25-year-olds in highly AI-exposed occupations sitting about 19% below where it would be had it kept pace with less-exposed peers, with no comparable gap for more experienced workers. The effect operates mainly through reduced hiring rather than layoffs. The episode holds this against the productivity finding to surface an uncomfortable pairing: AI helps junior workers most, and junior workers are the ones being hired less.
Sources & References
GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models, OpenAI — https://openai.com/index/gpts-are-gpts/
Gen-AI: Artificial Intelligence and the Future of Work, IMF — https://www.imf.org/en/Publications/Staff-Discussion-Notes/Issues/2024/01/14/Gen-AI-Artificial-Intelligence-and-the-Future-of-Work-542379
Gartner Predicts Agentic AI Will Autonomously Resolve 80% of Common Customer Service Issues by 2029 — https://www.gartner.com/en/newsroom/press-releases/2025-03-05-gartner-predicts-agentic-ai-will-autonomously-resolve-80-percent-of-common-customer-service-issues-without-human-intervention-by-20290
Gartner Survey Finds 85% of Service and Support Leaders are Expanding Human Agent Responsibilities — https://www.gartner.com/en/newsroom/press-releases/2026-04-28-gartner-survey-finds-eighty-five-percent-of-service-and-support-leaders-are-expanding-human-agent-responsibilities-despite-expectations-of-mass-ai-layoffs
Generative AI at Work, Brynjolfsson, Li & Raymond, NBER — https://www.nber.org/system/files/working_papers/w31161/w31161.pdf
The Future of Jobs Report 2025, World Economic Forum — https://www.weforum.org/publications/the-future-of-jobs-report-2025/
Canaries in the Coal Mine? Stanford Digital Economy Lab — https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/
Challenger Report: Layoffs Fall, Hiring Picks Up, AI Leads for Fifth Straight Month — https://www.challengergray.com/blog/challenger-report-layoffs-fall-hiring-picks-up-ai-leads-for-fifth-straight-month/
Voice narration is AI-generated.
Part two of a two-episode look at the money and politics behind AI data centers. After part one covered how projects move through planning under shell companies, code names, and non-disclosure agreements, this episode covers what happens when communities find out and push back — and the numbers show they are pushing back at a scale that surprised the industry.
Data Center Watch counted at least 75 projects worth roughly $130 billion blocked or delayed in the first three months of 2026 alone, matching the total for all of 2025. Active opposition groups more than doubled over the previous quarter and now operate in 49 states. More than 300 state data-center bills were filed in the first six weeks of the year, with statewide moratorium proposals introduced in 14 states, and Maine came within a single House vote of banning new data centers outright. In July 2026, New York became the first state in the country to impose a statewide pause on new hyperscale data centers while regulators write tougher standards. A national Reuters/Ipsos poll found a majority of Americans opposed having an AI data center built in their own community.
The episode uses Pacific, Missouri as a close-up case: hundreds of residents packed meetings over a proposed $16 billion campus, a February zoning meeting ended within minutes when the developer tabled its request, and months later — after the city passed a one-year moratorium — the developer withdrew the application.
On what drives the opposition, electricity costs rank near the top. One analysis of utility filings identified more than $4 billion in transmission projects approved in 2024 solely to connect data centers in parts of the PJM grid, with most of those costs passed to utility customers. Water, noise, diesel backup generators, land use, and tax incentives worth more than $1 billion a year in some states add further layers.
The economic case on the other side gets equal treatment. Virginia's legislative research agency, JLARC, estimates the industry supports about 74,000 jobs and $9.1 billion in annual state GDP — but the same report found most of that benefit comes from construction rather than ongoing operations. The contrast is stark: a typical build puts roughly 1,500 workers on site at peak, while a typical 250,000-square-foot facility runs on about 50 full-time workers once open, roughly half of them contractors. Supporters also warn that slowing construction too much could cost the United States ground in a technology race with economic and national-security stakes.
The episode closes on how the fight is shifting from yes-or-no toward the terms of the deal, covering federal power-cost pledges, Microsoft's commitments on utility rates and tax incentives, and the special rates, water fees, disclosure requirements, and community benefit agreements governments are now testing.
Sources & References
Q1 2026 Report, Data Center Watch — https://www.datacenterwatch.org/q1-2026
First Statewide Moratorium on New Hyperscale Data Centers, Office of Governor Kathy Hochul — https://www.governor.ny.gov/news/first-statewide-moratorium-new-hyperscale-data-centers-launched-governor-kathy-hochul
Data Centers in Virginia, JLARC — https://jlarc.virginia.gov/landing-2024-data-centers-in-virginia.asp
Data center opponents have blocked or delayed projects worth nearly $130 billion in 2026, NBC News — https://www.nbcnews.com/tech/tech-news/data-center-opposition-sharply-rising-2026-study-finds-rcna349728
Voice narration is AI-generated.
Amazon, Microsoft, Google, and Meta are expected to spend more than $700 billion in capital expenditures this year, much of it on AI infrastructure. Less visible is how those projects begin — often through a separate limited-liability company, a real-estate developer, or a project code name rather than a name residents would recognize. This is part one of two on the money and politics behind AI data centers, focused on what the public knows before the deals are underway.
The episode walks through documented cases. In Arkansas, Google built data centers in both Little Rock and West Memphis, working through intermediary firms in each case before confirming its own involvement, with reporting later finding the same person managing the intermediaries for both projects. In North Carolina, an Amazon project worth roughly $10 billion moved through local proceedings as Project Blue Marlin. In Memphis, some city council members and community leaders said they learned Elon Musk's xAI was building its supercomputer facility only the night before or the day of the public announcement.
On non-disclosure agreements, an NBC News review of more than 30 data-center proposals across 14 states found that in a majority of cases local officials had signed NDAs while also dealing with companies obscuring the identity of the ultimate developer, and five elected officials in different counties said their agreements barred them from sharing information with constituents. Virginia researcher Eric Bonds used public-records requests and found NDAs in 25 of 31 Virginia localities with existing or proposed data centers — roughly 80%. Some agreements reported around the country have even required officials to keep the existence of the NDA itself confidential.
Both sides of the argument get covered. Developers and economic-development officials cite legitimate reasons: weighing multiple sites, negotiating land purchases, protecting competitive plans, and the fact that details surface eventually when zoning, permits, or incentives require public approval. Critics focus on timing — land optioned, utilities consulted, incentives discussed, and designs advanced before residents know the company or the scale, so that by the first public hearing substantial planning and money may already be committed.
One major company has changed course. On March 18, 2026, Microsoft announced it would stop entering into NDAs with local governments for data-center projects worldwide and terminate those already active, saying transparency with communities mattered more, while still using confidentiality for private land transactions and protecting trade secrets and facility designs.
The episode is careful not to overstate the case: public hearings, zoning rules, environmental reviews, and permitting still apply depending on project and location.
Sources & References
How Big Tech uses NDAs to hide AI data center details from Americans, NBC News — https://www.nbcnews.com/tech/tech-news/data-center-ai-google-amazon-nda-non-disclosure-agreement-colossus-rcna236423
Microsoft nixes NDAs with local governments worldwide when deploying data centers, GeekWire — https://www.geekwire.com/2026/microsoft-nixes-ndas-with-local-governments-worldwide-when-deploying-data-centers/
Inside the Secretive Deal for a $10 Billion Data Center in Rural North Carolina, Inside Climate News — https://insideclimatenews.org/news/04082026/north-carolina-amazon-data-center-secret-deal/
Arkansas Explained: Understanding the data center boom and debate, Arkansas Advocate — https://arkansasadvocate.com/2026/05/22/arkansas-explained-understanding-the-data-center-boom-and-debate/
Voice narration is AI-generated.
The "cloud" turns out to be very physical. This episode breaks down what AI data centers consume — electricity and water — and why the numbers you see quoted online vary so wildly that some of them contradict each other outright.
On electricity, a single text prompt uses less than most people assume: Google measured its median Gemini text prompt at 0.24 watt-hours, roughly the energy of watching nine seconds of television. Heavier workloads are a different story, with long reasoning queries, multi-step agents, image generation, and especially video demanding far more computation — a long reasoning query can use more than ten times the energy of a standard one.
Scale is where it adds up. Data centers consumed about 1.5% of the world's electricity in 2025, and the International Energy Agency expects that to roughly double by 2030, reaching around 945 terawatt-hours. That global figure sounds modest until you look locally: Ireland's data centers consumed 23% of the country's metered electricity in 2025, and in the United States nearly half of data center capacity sits in a handful of regional clusters. The episode also explains why generating more power isn't the whole bottleneck — new data centers go up faster than the transmission lines, transformers, grid connections, and power plants needed to support them, making access to electricity one of the biggest constraints on new AI infrastructure.
On water, the episode clarifies a common misunderstanding: AI doesn't inherently need water. Chips need electricity and cooling, and many large facilities use water-based or evaporative cooling where some water is lost to evaporation, while others use air cooling, closed-loop liquid systems, or recycled water. A second footprint is easy to miss — generating electricity itself consumes water at the power plant.
That explains the contradictory per-prompt figures. One older academic estimate suggested GPT-3 could consume roughly a standard bottle of water per 20 to 50 prompts, counting both cooling and electricity-related water, while Google measured direct data center water use for a median Gemini prompt at about a quarter of a milliliter — roughly five drops. Both can be accurate, because they measure different models, years, hardware, and parts of the footprint.
The episode closes on what matters more: total volume and location. Google's environmental reporting implies companywide freshwater consumption of roughly 10 billion gallons in 2025, though companies don't cleanly separate AI workloads from everything else. Water use poses a very different problem in a cool, water-rich region than in a hot, drought-prone one. And the central tension is that individual tasks keep getting more efficient while total demand climbs anyway, as usage, model capability, and infrastructure all expand at once.
Sources & References
Measuring the environmental impact of AI inference, Google Cloud — https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference
Energy demand from AI, International Energy Agency — https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai
Data Centres Metered Electricity Consumption 2025, Central Statistics Office Ireland — https://www.cso.ie/en/releasesandpublications/ep/p-dcmec/datacentresmeteredelectricityconsumption2025/keyfindings/
Making AI Less "Thirsty", Li et al., arXiv:2304.03271 — https://arxiv.org/abs/2304.03271
Voice narration is AI-generated.
A bubble is when prices rise far ahead of proven economic value, pushed by excitement and expectation, until reality catches up. So the question about AI was never whether the technology works — it's whether the prices and spending around it have moved ahead of the returns it can produce. This episode lays out both sides of that argument using current figures rather than vibes.
On the scale of the spending: Gartner expects worldwide AI spending to total $2.5 trillion in 2026, growing 47% in a single year, with a large share going to data centers, servers, and chips. The largest technology companies are expected to spend roughly $800 billion on capital expenditures in 2026 alone. Sam Altman has said OpenAI carries roughly $1.4 trillion in data center commitments against about $20 billion in annual recurring revenue — a gap that sits at the center of the skeptical case.
The episode also examines the overlapping deals that draw criticism: OpenAI's reported $300 billion agreement with Oracle, a partnership with AMD to deploy six gigawatts of chips carrying warrants that could give OpenAI a significant stake in the chipmaker, and an announced Nvidia plan that could have involved up to $100 billion of investment before Nvidia said it no longer expected that figure. Critics argue such arrangements blur independent customer demand and demand underwritten by suppliers; supporters call them a rational way to lock up scarce capacity.
The counterargument gets equal weight. Today's AI leaders are large, profitable businesses — Nvidia posted $31.8 billion in non-GAAP net income in a single quarter. Valuations sit well below dot-com extremes: Cisco traded north of 150 times forward earnings at its peak, while Nvidia has been trading in the low thirties on trailing earnings.
One recent quarter captures the strangeness of the moment. Nvidia reported $81.6 billion in revenue, ahead of expectations, with data center revenue up 92% year over year — and the stock fell anyway, the fourth straight quarter where Nvidia beat and shares still dropped. When investors already expect extraordinary growth, extraordinary results stop being enough.
The episode closes on Cisco as the cautionary case: it helped build the internet's backbone, peaked near $560 billion in market value, sits closer to $200 billion today, and has never returned to that high more than 25 years later. A technology can transform the world while the prices around it still overheat. What decides the verdict in hindsight is whether the infrastructure produces enough economic return to justify the spending. Nothing here is investment advice.
Sources & References
Gartner Says Worldwide AI Spending Will Total $2.5 Trillion in 2026 — https://www.gartner.com/en/newsroom/press-releases/2026-1-15-gartner-says-worldwide-ai-spending-will-total-2-point-5-trillion-dollars-in-2026
Sam Altman says OpenAI has $20B ARR and about $1.4 trillion in data center commitments — https://finance.yahoo.com/news/sam-altman-says-openai-20b-211806316.html
NVIDIA Q3 FY 2026 Earnings: Record Data Center Revenue, Futurum — https://futurumgroup.com/insights/nvidia-q3-fy-2026-record-data-center-revenue-higher-q4-guide/
Nvidia Earnings Beat Estimates as Data Center Revenue Surges 92%, Traderverse — https://traderverse.io/blog/nvidia-earnings-data-center-revenue-surges-92-as-stock-slips
From Cisco to Nvidia: Lessons From Past Tech Bubbles, Creative Planning — https://creativeplanning.com/insights/investment/cisco-nvidia-lessons-from-past/
Voice narration is AI-generated.
In June 2025, OpenAI's Sam Altman said Meta was offering his staff $100 million signing bonuses. The number traveled everywhere — and then came apart. Meta's CTO Andrew Bosworth said the figure described multi-year total compensation for a few very senior leadership roles rather than cash on day one, and Lucas Beyer, a researcher who did leave OpenAI for Meta, said plainly on social media that no $100 million sign-on had happened. This episode uses that gap between headline and offer letter to explain what AI engineers are genuinely paid and why.
The broad market has measurable numbers behind it. PwC's 2026 Global AI Jobs Barometer, drawn from more than a billion job advertisements across 27 countries, found that workers with AI skills command a 62% wage premium over comparable roles without them, up from 57% the previous year. Demand is climbing alongside it: postings requiring AI skills grew 69%, against 9% for the overall job market. That premium is broad and documented, and nothing like the headline figures.
Above that sits a much smaller tier of researchers whose expertise is considered critical to building frontier models. For scale, Meta's named executive officers earn roughly $20 to $24 million a year in total compensation — the neighborhood of the ceiling at the very top of a large public company. One documented offer to an AI researcher that did surface publicly was $18 million, which the researcher declined.
The episode covers three forces behind top-tier pay: scarcity, since very few people have deep experience training frontier models while several well-funded companies pursue the same handful; leverage, since one person's work on a training method or safety technique can shape a product used by millions; and simple spending capacity, since companies already committing extraordinary sums to chips and data centers treat talent as another line in the same race.
It also explains why package structure matters as much as size. Large offers combine salary, bonuses, and equity vesting over several years — equity that is sellable stock at a public company but often much harder to convert to cash at a private lab, making headline value and spendable money two different figures. That structure is why two people sharing the title "AI engineer" can occupy entirely different financial universes.
The closing section weighs whether the pay is justified, presenting the case that scarcity and leverage are genuine against the case that bidding has been inflated by competition and fear of falling behind.
Sources & References
PwC 2026 Global AI Jobs Barometer — https://www.pwc.com/gx/en/news-room/press-releases/2026/pwc-2026-ai-jobs-barometer.html
Meta is offering multimillion-dollar pay for AI researchers, but not $100M 'signing bonuses', TechCrunch — https://techcrunch.com/2025/06/27/meta-is-offering-multimillion-dollar-pay-for-ai-researchers-but-not-100m-signing-bonuses/
Sam Altman says Meta is offering $100 million to poach OpenAI employees, CNN Business — https://www.cnn.com/2025/06/18/tech/meta-openai-sam-altman-100-million
Voice narration is AI-generated.
A language model knows what it learned during training and whatever you hand it in the current conversation. It doesn't know what happened yesterday, what's in your company's internal handbook, or what's inside a document you never gave it — and retraining a model every time a document changes isn't practical. RAG, short for retrieval-augmented generation, is the pattern built to close that gap, and this episode explains how it works and where it breaks down.
The idea is straightforward: before the model writes anything, go find the relevant information and hand it over, then ask the model to answer using that material. Retrieval, then generation. The term traces back to a 2020 research paper from Patrick Lewis and colleagues, which paired a trained model with a searchable index it could pull from.
The episode walks the mechanism step by step. You start with a knowledge base — a collection of documents. When a question arrives, a retrieval step searches that collection and pulls the passages most likely to be relevant. That material goes into the model's context alongside the question, and the model generates its answer from what it was just handed rather than from training alone. It connects this directly back to the context window from episode one: retrieval is the decision about which pieces of information are worth spending that limited space on. Good retrieval also does more than keyword matching — it can search by meaning, so a question phrased one way still finds a passage that says the same thing in different words.
A worked example makes it concrete. A customer asks a company's support assistant about the refund policy. The model was never trained on that company's handbook, so the system retrieves the handbook section on refunds, places it into the model's context, and the model answers from that specific material — the difference between answering from memory and answering with the reference page open in front of you.
The episode closes on why RAG became so widely used and where it fails. Answers stay current because you update source material instead of retraining; answers are grounded in identifiable sources the system can point back to; and a model can work with private or specialized information without that information ever being trained in. The catch is that RAG depends heavily on retrieval quality — fetch the wrong passage or miss the right one and the answer suffers. Document preparation, result ranking, how much context gets included, and how the model is prompted all shape the outcome too. Handing a model the correct reference helps enormously, but it doesn't make the model infallible.
Sources & References
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, Lewis et al., arXiv:2005.11401 — https://arxiv.org/abs/2005.11401
Voice narration is AI-generated.
Hook an AI assistant up to your files, your calendar, and a database at work, and someone has to build all three connections by hand. Add a fourth tool and you wire it up again; switch assistants and much of that work starts over. MCP, the Model Context Protocol, exists to end that one-off plumbing. This episode explains what MCP actually is, how it's structured, and what a company is really claiming when it says its app supports MCP.
MCP is an open standard introduced by Anthropic in 2024, and it is no longer Anthropic's alone. In December 2025, Anthropic donated the protocol to the Agentic AI Foundation, a vendor-neutral project under the Linux Foundation, co-founded alongside Block and OpenAI. The project's own analogy for itself is a USB-C port for AI applications: rather than every application inventing a private way to reach every tool and data source, MCP provides one common interface.
The episode walks through the architecture. The AI application is the host. Inside it, an MCP client handles standardized communication with one or more MCP servers, and an MCP server is a program that exposes capabilities to the application — access to GitHub, a database, a file system, or another service. Because client and server follow the same protocol, the application can discover what a server offers and interact with it through a consistent pattern.
Servers can expose three kinds of capabilities: tools, which are actions the model can invoke like running a search or writing a file; resources, which are data the application can read such as documents or database records; and prompts, which are reusable templates for common tasks. A server might expose one, several, or all three.
The episode is careful not to oversell interoperability. A standardized connection pattern is not the same as universal compatibility — protocol versions, supported capabilities, authentication, permissions, and transport all still matter, and not every client works with every server in every situation. It also connects MCP back to the agent from episode 2: the agent decides what to do, while MCP is the common doorway it uses to reach outside tools and data.
It closes on the security question. A doorway to real files, accounts, and services raises genuine questions about what an application can reach, what it can change, and what credentials it uses. A standard makes connecting easier, but deciding what should be connected still takes judgment.
Sources & References
Introduction, Model Context Protocol documentation — https://modelcontextprotocol.io/docs/getting-started/intro
Architecture overview, Model Context Protocol documentation — https://modelcontextprotocol.io/docs/2026-07-28/learn/architecture
MCP joins the Agentic AI Foundation, Model Context Protocol Blog — https://blog.modelcontextprotocol.io/posts/2025-12-09-mcp-joins-agentic-ai-foundation/
Linux Foundation Announces the Formation of the Agentic AI Foundation — https://www.linuxfoundation.org/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation
Voice narration is AI-generated.
AI agent. Agentic. The words are everywhere, and they sometimes sound like a buzzword and sometimes like a specific piece of technology. This episode makes the distinction concrete: what separates an agent from an ordinary language model interaction, what an agent is built from, and what question to ask when a product claims to be one.
A basic model interaction is simple — you send a prompt, the model generates a response, the request ends. An agent goes further, taking actions toward a goal rather than only producing an answer: searching the web, calling a program, querying a database, editing a file, or sending a request to another service, then using each result to decide the next move. Agents run in a loop, where the model receives the current context, chooses whether to answer or use a tool, the system executes that action, and the result feeds into the next step, repeating until the task is done, the system hits a limit, or it needs a person's approval.
The episode uses Anthropic's own distinction between workflows and agents to draw the line on autonomy. In a workflow, software follows a path designed ahead of time. In an agent, the model directs its own process, deciding which steps to take and which tools to use as it works toward the goal.
It then breaks down the ingredients: the model that interprets the situation and picks the next move, instructions defining the task and its boundaries, tools the system is permitted to use, the loop that feeds results back in, and working context to track the current run. Longer-term memory that survives between runs is optional rather than required. A worked example contrasts a language model with no tools, which can explain how to look for a flight but cannot check live inventory, against an agent with access to a flight-search service, which can work out the date, search current options, compare results, and return a short list.
On risk, the episode draws directly on Anthropic's published guidance: autonomy raises costs and creates the potential for compounding errors, since an agent that misreads a result early will build later actions on that mistake. Anthropic recommends extensive testing in sandboxed environments, appropriate guardrails, and pausing for human feedback at checkpoints. The closing point is that autonomy is a spectrum rather than a fixed design — some agents ask permission before consequential actions, others receive a goal and wide latitude to reach it.
Sources & References
Building effective agents, Anthropic — https://www.anthropic.com/engineering/building-effective-agents
Voice narration is AI-generated.
If you've spent any time around AI, you've seen prices, limits, and context windows measured in tokens — which makes a token sound like a technical word for a word. It isn't quite that simple. This episode explains what a token is, how a tokenizer decides where one ends and the next begins, and why the answer shapes what you can fit into a model and what you pay to use it.
As the first episode of Learn AI in Bits, it also introduces the show: AI explained from the basics through more advanced ground, plus the news worth knowing, in about five minutes an episode, with longer deep dives when a subject needs the room. The show is made by a former software engineer, now a technical project lead in the AI space, and uses AI for research, preparation, plain-language translation, and narration, with every episode reviewed before release.
A token is a unit of text a language model processes: sometimes a whole word, sometimes part of one, sometimes a punctuation mark or a chunk of text that includes a space. Before a model touches your sentence, a tokenizer converts it into a sequence of tokens, each mapped to a number called a token ID, which the model then turns into learned numerical representations known as embeddings. Different model families use different tokenizers — OpenAI's tiktoken library uses byte pair encoding, or BPE, which builds tokens from byte patterns that occur frequently, representing common words compactly while splitting unusual text into smaller pieces.
The episode walks through a worked example drawn directly from OpenAI's own documentation, showing how the single word "red" produces three different tokens depending on context: lowercase with a leading space as it appears mid-sentence, capitalized with a leading space, and capitalized with no space as it appears at the start of a sentence. Same three letters, three distinct tokens, because the tokenizer treats spacing and capitalization as part of the pattern it matches.
From there it covers OpenAI's published rule of thumb for English — roughly four characters or about three-quarters of a word per token — along with the caveat that exact counts depend on the tokenizer, the model encoding, and the text itself, and that tokenization varies across languages.
The practical stakes come in three parts. Tokens define the context window, the maximum amount of tokenized information a model can work with in one request, a budget that covers both input and generated output. Tokens are also the billing unit when using a model through an API, where input and output tokens are priced separately and some models report cached or reasoning tokens as well, meaning two prompts with identical word counts can produce different costs. And tokens are the mechanism behind generation itself: a model produces text one token at a time, calculating probabilities across possible next tokens based on everything processed so far.
Sources & References
What are tokens and how to count them?, OpenAI Help Center — https://help.openai.com/en/articles/4936856-what-are-tokens-and-how-to-count-them
tiktoken, OpenAI — https://github.com/openai/tiktoken
Voice narration is AI-generated.
From the publisher's feed