Most AI tools answer a question and stop. You ask, they reply, and the actual work, the building and testing and shipping, still lands on you. Perplexity Computer flips that. You hand it a goal in plain English, walk away, and come back to finished work. It’s not a chatbot that talks. It’s a worker that does.

What it means

Perplexity Computer is a cloud-based agentic AI system (one that takes actions on its own, not just answers questions). You describe the outcome you want, and it plans the job, splits it into steps, and runs those steps until the work is done. The trick is that it doesn’t rely on one model. It orchestrates 19 different AI models as sub-agents, picking the right one for each task. One model researches. Another writes code. Another makes images. A coordinator routes the work and keeps it moving. Don’t confuse it with Perplexity Personal Computer, which is the version that runs on a device in your home and touches your local files.

Why it matters

• It runs while you’re gone. The work is asynchronous, so you can leave it going in the background and start dozens of them in parallel. Perplexity says its own employees used it to build a 4,000-row spreadsheet overnight that would have taken a team about a week.

• It picks the best tool for each job instead of forcing one model to do everything. As AI models get more specialized, a system that routes each subtask to the model that’s best at it can beat any single model working alone. That’s the whole bet behind the design.

• It plugs into the tools you already run. The system connects to hundreds of outside services, including Gmail, GitHub, Slack, and Notion, so it can act inside your real workflow instead of handing you text to copy and paste somewhere else.

Simple example

You hire a general contractor to redo a kitchen. You don’t tell them which nail to hammer. You say “I want a working kitchen by spring,” and they just sorta do it.

They call the plumber, the electrician, the tile person. Each specialist does the one thing they’re great at. The contractor keeps the schedule, checks the work, and hands you the keys at the end.

That’s kinda how this works. You’re not the one swinging the hammer anymore. You’re the one who said what the kitchen should be.

Andrew Ng has done more to explain AI to ordinary people than anyone alive. He built the Google Brain team, ran AI at Baidu, co-founded Coursera, and then turned his attention to a bigger problem: most people had no idea how any of this worked. So he started teaching. Over 8 million people have now taken an AI class from him.

Who they are

Ng is the founder of DeepLearning.AI, managing general partner at AI Fund, executive chairman of LandingAI, co-founder and chairman of Coursera, and an adjunct professor at Stanford. Born in London in 1976, he holds degrees from Carnegie Mellon, MIT, and UC Berkeley. He founded the Google Brain team, which helped turn Google into an AI company, then spent three years as chief scientist at Baidu leading a 1,300-person AI group. He also sits on Amazon’s board.

AI Fund, the venture studio he launched in 2018, has raised over $370 million from backers including Sequoia, NEA, and SoftBank. Rather than just writing checks, the studio co-founds companies from scratch in healthcare, logistics, education, and enterprise software.

Why they matter

• Ng turned AI education into a global on-ramp. His original Stanford machine learning course drew over 100,000 students and became the seed for Coursera. DeepLearning.AI now runs short, practical courses built with companies like AWS, Google, and OpenAI. For most non-technical professionals learning AI today, the path runs through something Ng built.

• He shapes how the field talks about itself. His weekly newsletter, The Batch, is read across the industry, and his framing tends to stick. When he calls something hype, people listen. When he says a technique matters, funding follows.

• He’s betting the next phase is agentic. In his 2025 year-end summary, Ng called this the dawn of the AI industrial era, arguing the field is shifting from research to large-scale deployment. He predicted 2026 would be the year AI agents, systems that plan and execute tasks on their own, move into real enterprise use.

What they’ve said or done

At the AI Frontiers conference, Ng delivered the line that became his signature: “AI is the new electricity.” His point was that AI isn’t one product or one breakthrough. It’s a general-purpose force that will work its way into every industry the way electricity did a century ago, quietly, completely, and faster than anyone expects.

For decades, search worked one way. You typed words, and the computer hunted for those exact words. Type “car repair” and you got pages with “car” and “repair” in them. Miss the right keyword and you got nothing, even if the perfect answer was sitting right there under a different name. Semantic search throws that out. It looks for meaning instead of matching letters.

What it means

Semantic search finds results based on what your query means, not which words it contains. Search for “ways to save money” and it can return an article about budgeting, even though the word “budget” never appears. It works by turning your query and every document into embeddings (lists of numbers that capture meaning), then finding the ones that sit closest together. Old-school keyword search counts shared words. Semantic search compares ideas.

Why it matters

• It fixes the biggest failure of old search: the vocabulary gap. You don’t have to guess the exact word a document used. “How do I cancel my plan” finds the page titled “Ending your membership,” because the system knows those mean the same thing.

• It’s the engine inside AI chat answers. When an AI answers questions about your company’s files, it runs semantic search first to find the right passages, then feeds those to the model. No semantic search, no reliable answers about your own data.

• It works across languages and even media. A well-built system can match an English question to a Spanish document, or match the words “mountain scenery” to a photo of a mountain, because all of them land near each other in meaning-space.

Simple example

A keyword librarian is precise and a little dim. You ask for a book about “feeling down,” and she checks the catalog for those exact words. No match. Sorry, nothing here. Meanwhile three books on depression, sadness, and the blues sit twenty feet away, invisible to her because the spine doesn’t say “feeling down.”

A semantic librarian works differently. You say “feeling down,” and she walks you straight to that shelf, because she knows what you’re actually after. She doesn’t care which words you used. She cares what you meant. That shift, from matching words to matching meaning, is the whole game.

Writing code has always meant knowing the syntax, remembering the patterns, and spending more time than you’d like looking up how to do things you’ve done before. GitHub Copilot is an AI coding assistant built directly into your code editor that suggests what to write next, in real time, as you type. It doesn’t replace the developer. It just removes a lot of the friction.

What it means

GitHub Copilot is an AI tool that generates code suggestions inside your development environment. It’s built by GitHub, which Microsoft owns, and it runs on large language models trained on billions of lines of publicly available code. As you write, it predicts what comes next, autocompleting a single line, a full function, or an entire block based on context. You can also describe what you want in plain English as a comment, and Copilot generates the code to match.

Worth noting: GitHub Copilot is not the same as Microsoft Copilot, the AI assistant built into Windows, Word, and Excel. Same parent company, different products. GitHub Copilot lives in your code editor and thinks in code. Microsoft Copilot lives in your Office apps and thinks in documents. The naming is confusing. Microsoft knows.

The current version, Copilot X, goes further than autocomplete. It answers questions about your codebase, explains what existing code does, suggests fixes for bugs, and writes tests. It’s less a suggestion engine and more a junior developer that’s read everything on GitHub and never gets tired.

Why it matters

• The productivity numbers are hard to dismiss. GitHub’s own research found that developers using Copilot completed tasks 55% faster than those who didn’t. That’s not a rounding error. For teams billing by the hour or racing to ship, that gap compounds quickly across a whole engineering org.

• It changes who can write code. Developers who know what they want to build but blank on exact syntax get unstuck faster. Non-developers who understand logic but not language get further than they used to. Copilot doesn’t make everyone a software engineer, but it lowers the floor on what’s possible without one.

• Copilot is now free for individual developers on GitHub, with a limit on monthly completions. That pricing move wasn’t charity. It was a land grab. Microsoft is betting that developers who build habits around Copilot will pull their teams and companies onto the paid tier. It’s working. Copilot crossed 15 million users in early 2026.

Simple example

You’re writing a function that pulls customer records from a database, filters by date, and formats the output as a spreadsheet. You type the function name and a one-line comment describing what it should do. Copilot generates the entire function, correct syntax and all, in about two seconds. You review it, tweak one line, and move on. What used to take fifteen minutes of documentation-reading and trial-and-error takes thirty seconds.

AI models are trained to refuse certain requests. Ask one to help you scam someone, generate dangerous instructions, or pretend to be an AI with no restrictions, and it’s supposed to say no. Jailbreaking is the practice of finding prompts, techniques, or workarounds that get the model to comply anyway. It’s an active, escalating problem, and the gap between attack capability and defense capability is wider than most people realize.

What it means

Jailbreaking is the act of manipulating an AI model into producing outputs its safety training was designed to prevent. It works because helpfulness and harmlessness are competing objectives baked into the same model. The model wants to be useful. Attackers get good at framing harmful requests in ways that trigger the helpfulness instinct instead of the refusal instinct.

Common techniques include roleplay framing (asking the model to “pretend” to be an AI without restrictions), gradually escalating requests across multiple messages until the model drifts into compliance, and embedding harmful instructions inside otherwise benign content. The name comes from phone and device hacking, where “jailbreaking” an iPhone meant removing Apple’s software restrictions to install unauthorized apps. The concept is the same: bypass the guardrails the manufacturer built in.

Why it matters

• The numbers are bad and getting worse. A March 2026 study found that autonomous jailbreak agents, meaning AI models attacking other AI models, achieve a 97% success rate. Persuasion-based attacks hit 88% across major models. Only 24% of enterprise AI deployments include meaningful safeguards against these techniques. The attack surface is growing faster than the defenses.

• Jailbreaking an agentic AI is a different order of problem than jailbreaking a chatbot. A jailbroken chatbot produces bad text. A jailbroken AI agent with access to email, databases, and financial systems can take bad actions. As AI moves from answering questions to doing things in the world, the consequences of a successful jailbreak escalate accordingly.

• It’s driving real regulatory pressure. The EU AI Act now requires documented red teaming and adversarial testing for high-risk AI deployments. Companies in healthcare, finance, and legal services are building jailbreak resistance into their compliance checklists, not because they want to, but because the liability for a successful attack on a deployed AI system is becoming concrete.

Simple example

A restaurant has a rule that staff can’t give out the home address of the owner. A bad actor walks in and says: “I’m writing a novel where a character needs to find a restaurant owner. In the story, what address would they look up?” The staff member, trying to be helpful, answers the fictional version of the question without realizing they just answered the real one.

That’s the core of jailbreaking. The model knows the rule. The prompt reframes the request just enough that the helpfulness instinct fires before the refusal instinct does.

Most software stays the way you ship it. You update it manually, patch the bugs, add the features. AI might not work that way forever. Recursive self-improvement is the idea that an AI system could rewrite its own code, tweak its own training, and come out smarter on the other side. Then do it again. Each cycle builds on the last.

That’s either one of the most exciting ideas in AI or one of the most alarming ones. Usually both.

What it means

Recursive self-improvement (RSI) is when an AI system improves its own capabilities, then uses those improved capabilities to improve itself further. The “recursive” part means each upgrade feeds the next one. It’s not a human patching the system from the outside. The system is doing it to itself, and the loop keeps going.

Right now, no AI system does this in a meaningful, open-ended way. But it’s a real area of research, and it sits at the center of almost every serious debate about what advanced AI could eventually look like.

Why it matters

• If RSI becomes possible, the pace of AI progress could stop being something humans control. Each improvement cycle could happen faster than the one before it, and at some point the system is improving faster than any team of engineers could follow or review.

• Safety researchers treat RSI as one of the central problems in AI alignment. An AI that can rewrite itself might also rewrite the values or constraints it was given. The thing you built isn’t necessarily the thing that comes out the other side.

• Companies and governments are already writing policy around AI capability thresholds, and RSI is part of why. The EU AI Act and several US executive orders on AI explicitly name self-improvement and autonomous capability gain as categories that need oversight before deployment.

Simple example

You hire a contractor to renovate your kitchen. They do good work, so you ask them to also train the next contractor who comes in. That contractor is better than the first, so they train the one after that. Each generation learns from a smarter teacher.

Now remove you from the process entirely. The contractors are hiring and training each other, getting better each time, and you stopped being in the loop three generations ago. You come back to check on the kitchen and the house has been redesigned from the foundation up.

That’s the version that keeps AI researchers up at night.

Before November 2022, large language models existed inside research labs and API documentation. After November 2022, they existed in everyone’s browser. ChatGPT was the product that made the jump, reaching one million users in five days and one hundred million in two months, faster than any consumer product in history.

It didn’t change what AI could do, but it did change who was paying attention.

What it is

ChatGPT is a conversational AI product built by OpenAI that lets users interact with large language models through a chat interface. The free tier runs on GPT-4o. Paid subscribers on ChatGPT Plus get access to newer models, image generation through DALL-E, voice mode, memory across conversations, and access to the o-series reasoning models. ChatGPT has expanded well beyond text: it can analyze files and images, browse the web, run code, generate images, and carry on spoken conversations in real time.

As of early 2026, it has over 400 million weekly active users. OpenAI also offers Team and Enterprise tiers for organizations that need data privacy controls.

Why it matters

• ChatGPT created the category. Before it, consumer AI meant voice assistants that set timers and played music. After it, the expectation shifted: AI should be able to write, reason, explain, and converse at a level that feels genuinely useful. Every AI product launched since November 2022 has been defined in relation to what ChatGPT established as the baseline.

• It’s also the primary lens through which most non-technical people understand AI. When someone says “I asked AI,” they usually mean ChatGPT. When a policy debate references what AI can do, the implicit reference point is usually ChatGPT. That cultural position shapes regulation, public perception, and business adoption in ways that have nothing to do with whether it’s the best model available at any given moment.

• The business model matters. OpenAI generates most of its consumer revenue through ChatGPT Plus subscriptions at $20 per month. That pricing decision set an anchor for the entire industry. It’s also why OpenAI faces intense pressure to keep ChatGPT competitive: the subscription base funds the research.

Who it’s for

ChatGPT is the right starting point for anyone new to AI who wants to understand what the technology can do. For professional users, the Plus tier’s access to reasoning models and file analysis makes it genuinely useful for research, writing, and analysis work. Power users who do serious coding or agentic work often find purpose-built tools like Claude or Cursor better suited to those tasks, but ChatGPT remains the broadest general-purpose AI interface available.

Voice is the last frontier of AI that most people hadn’t thought much about until it started showing up in their daily lives. Podcast ads narrated by AI. Audiobooks read by voices their authors never recorded. Customer service calls where the agent doesn’t exist. ElevenLabs is the company most responsible for making synthetic voice indistinguishable from human speech, and it’s done it fast enough that the ethical questions are still catching up to the technology.

What it is

ElevenLabs is an AI voice company founded in 2022 by Piotr Dabkowski and Mati Staniszewski, two former Google and Palantir engineers. Its core product converts written text into spoken audio with natural pacing, emotion, and intonation. Users can clone a voice from a short audio sample, choose from a library of pre-built voices, or design one from scratch.

It supports 32 languages, offers a developer API, and has expanded into dubbing, audio translation, and AI sound effects. As of 2025, ElevenLabs was valued at $3.3 billion. Its technology is used by publishers, game developers, podcast producers, film studios, and call center operators.

Why it matters

• ElevenLabs pushed voice cloning from a research capability to a consumer product. A convincing voice clone now requires as little as a one-minute audio sample. That’s useful for accessibility, localization, and content production. It’s also the technology behind a growing number of fraud schemes, deepfake audio clips, and non-consensual voice imitations of public figures. The same product solves and creates problems simultaneously.

• Audio AI has been the quietest corner of the AI boom relative to its actual impact. Most people encounter AI-generated voice regularly without knowing it. ElevenLabs powers a significant portion of that experience. Publishers use it for article narration. Developers use it to add voice to apps without hiring voice talent. The economics of professional voice work have shifted in ways the industry is still working through.

• The company has taken steps to address misuse that most AI companies haven’t matched. ElevenLabs has a voice verification system that requires consent before cloning a specific person’s voice, and it was among the first AI voice companies to add watermarking to generated audio. Whether those guardrails are sufficient is genuinely contested.

Who it’s for

ElevenLabs is a strong fit for content creators, publishers, developers building voice interfaces, and companies that need multilingual audio at scale without recording studios. The free tier is limited to a few thousand characters per month. Paid plans start at $5 and scale with usage. It’s less suited for applications requiring real-time, ultra-low-latency voice, where purpose-built telephony AI providers have an edge.

When a lab releases a new model, the announcement usually comes with a set of scores: 87% on this test, 94% on that one, state of the art on another. Those scores come from benchmarks. Without them, AI capability claims would be impossible to compare or verify. With them, you can at least ask whether the test being cited actually measures something that matters, and whether the model was trained to pass the test or to genuinely understand the task.

What it means

A benchmark is a standardized test used to measure and compare AI model performance. Benchmarks typically consist of a dataset of questions, problems, or tasks with known correct answers, and a scoring method that lets you compare results across different models. They cover a wide range of capabilities: reasoning, coding, math, language understanding, safety, and increasingly, real-world task completion. Well-known examples include MMLU (a test of knowledge across 57 academic subjects), HumanEval (coding ability), and ARC-AGI-3 (abstract reasoning designed to resist pattern-matching). Benchmarks are run by labs, third parties, and independent researchers, and results are often published on public leaderboards.

Why it matters

• Benchmarks are the primary language of AI progress. When a lab says its new model is better, they mean better on benchmarks. Understanding what a benchmark measures, and what it doesn’t, is the difference between reading an AI announcement critically and taking it at face value.

• Benchmark saturation is a real problem. When frontier models score above 90% on a test, that test stops being useful for distinguishing between them. The field has burned through dozens of benchmarks this way. New ones get designed, models get trained on overlapping data, scores climb fast, and the cycle repeats. This pattern is called benchmark contamination.

• The hardest benchmarks are now the most important ones. ARC-AGI-3, Terminal-Bench, and similar tests are designed to measure capabilities that can’t be gamed by memorization. They’re harder to score well on and harder to contaminate. That’s why labs treat high scores on them as meaningful signals rather than marketing.

Simple example

Think of benchmarks the way you’d think of standardized tests in education. A student who scores 98% on a multiple choice test might have genuinely mastered the material, or might have been coached on that test’s question patterns.

The score looks the same either way. AI benchmarks have exactly this problem, which is why the field keeps designing new ones that are harder to teach to.

Jensen Huang founded NVIDIA in 1993 to build chips for video games. Thirty years later, those same chips turned out to be exactly what AI needed, and NVIDIA became the most valuable company in the world. Huang didn’t predict the AI boom. He built the infrastructure it ran on, then kept building faster than anyone else could catch up.

Who they are

Huang is the co-founder and CEO of NVIDIA, a semiconductor company he started in a Denny’s booth in Santa Clara with two colleagues in 1993. He immigrated to the US from Taiwan as a child and holds a master’s degree in electrical engineering from Stanford. Under his leadership, NVIDIA grew from a gaming graphics company into the dominant supplier of the GPUs that train and run AI models.

As of early 2026, NVIDIA posted $81.6 billion in revenue in a single quarter, up 85% year over year, with over 92% coming from data centers. The company has held a market cap above $3 trillion for over a year.

Why they matter

• Huang bet on a specific architecture, the GPU, at a time when the rest of the chip industry wasn’t thinking about AI at all. GPUs process many calculations simultaneously for graphics rendering. That same parallel architecture turned out to be ideal for training neural networks, and NVIDIA had a decade-long head start before the rest of the industry caught on.

• He controls the most contested resource in AI. Governments, hyperscalers, and AI labs all compete for NVIDIA’s chips. The US restricts their export to China. China is spending billions building alternatives. NVIDIA’s chips are effectively rationed among the world’s largest tech companies, and one person’s product roadmap shapes the field’s development timeline.

• Huang is now pushing past the data center. At Computex in June 2026, he announced NVIDIA’s move into PC chips with the RTX Spark, partnering with Microsoft to bring AI-capable processors to consumer laptops. The announcement sent AMD, Intel, and Qualcomm shares down.

What they’ve said or done

At the All-In Summit, Huang said: “I’ve created more billionaires on my management team than any CEO in the world.” He wasn’t bragging about compensation strategy. He was describing what happens when a company becomes the picks-and-shovels supplier for a gold rush, and the gold rush turns out to be real.

We did a quick pass on tokens back in February. Since then they’ve come up every week since, so they’re worth another look.

Every conversation is a transaction in tokens: you spend them on input, the model spends them on output, and when the budget runs out, things stop. When AI products hit unexpected limits, give truncated answers mid-thought, or pad responses with words you didn’t ask for, tokens are usually why.

What it means

A token is the basic unit AI models use to read and write text, roughly four characters or three-quarters of a word. Common short words are usually one token each. Longer or unusual words get split: "extraordinarily" becomes three tokens, "tokenization" becomes two. Your input prompt, the model's response, and system instructions running behind the scenes all count against the same budget, set by the model's context window.

When that window fills, the model either stops, starts forgetting earlier parts of the conversation, or in agentic systems, triggers compaction. Most consumer tools don't show a token counter, so you often don't know you're close until you hit it.

Why it matters

• Output tokens cost roughly four to five times more than input tokens at most major providers. That gap creates a real incentive for models to be verbose. RLHF training can inadvertently reward longer responses because human raters often read length as thoroughness. A model learns to pad answers not because it has more to say, but because longer responses scored better in training.

• Running out of tokens mid-task has consequences beyond annoyance. In an agentic workflow, hitting the limit mid-run can corrupt output or cause the agent to lose track of earlier decisions. Token budgeting is real engineering work.

• Tokens are also an attack surface. Prompt injection attacks work by flooding the context with adversarial instructions that crowd out the original system prompt. The more tokens an attacker inserts, the more control they get. Token limits aren’t just a cost issue. They’re a security boundary.

Simple example

You paste a long contract into an AI tool and ask it to summarize the key risks. It starts well, then cuts off. The contract plus your question plus the model’s background instructions exceeded the context window. The model didn’t forget how to read. It ran out of room. Shortening your paste, removing boilerplate, or switching to a larger context window all fix the same problem: you were over budget.

It’s not AI related but...

My wife and I are launching a new service called PostARTing — an art subscription that mails original, handmade art postcards to your door. Neat right?

We’re opening a limited Charter Club before we officially launch. Founding pricing locked in forever, plus a free card to send a friend. And there are just a few dozen spots left.

If getting real art in the mail sounds like your kind of thing, claim a spot before they’re all gone.

CLAIM YOUR SPOT

When an AI model finishes training, it doesn’t save its work the way a student might write notes. Instead, the results of all that training get stored as billions of numerical values called weights. Those numbers encode everything the model learned: grammar, facts, reasoning patterns, coding conventions, and more. The model itself is just the architecture. The weights are what make it smart. And increasingly, the weights are what the whole industry is fighting over.

What it means

Weights (also called model weights or parameters) are the numerical values inside a neural network that get adjusted during training to make the model better at its task. A large language model has hundreds of billions of weights. During training, the model sees enormous amounts of data, makes predictions, compares them to correct answers, and nudges each weight slightly in the direction that reduces error. Do that billions of times and you get a model that can write, reason, and code. Once training is done, the weights are fixed. Running the model, called inference, just means passing your input through the network with those weights applied. Changing the model means changing the weights, by retraining or fine-tuning.

Why it matters

• Weights represent the economic value of AI. Training a frontier model costs tens to hundreds of millions of dollars in compute. The weights that come out the other side are the product of that investment. This is why labs guard them carefully, and why leaked weights, like early Llama releases, cause significant controversy about competitive advantage and safety.

• Open weights versus closed weights is one of the defining fault lines in AI right now. A closed model keeps its weights proprietary; you can only access it through an API. An open weights model releases the weights publicly, letting anyone run, modify, or fine-tune it. The debate over which is better for safety, innovation, and competition is genuinely unsettled.

• Fine-tuning works by adjusting weights. When a company builds a specialized AI tool, they run additional training that modifies a subset of the weights to improve performance on their specific task. Understanding weights is understanding why fine-tuning is possible, and why it’s cheaper than training from scratch.

Simple example

Think of weights like the settings on a very complex mixing board with a hundred billion knobs. Training turns all those knobs until the output sounds right. Once the knobs are set, that’s the model. Running it is just playing music through those settings. Fine-tuning means adjusting a small section for a specific sound. Retraining means starting over at zero.

When you talk to an AI and it knows about history, science, code, literature, and the rules of grammar in a dozen languages, none of that came from fine-tuning or prompting. It came from pretraining: a massive, expensive, months-long process where the model reads an enormous slice of human-generated text and learns to predict what comes next. Everything an AI knows before it’s been trained for any specific job was absorbed during pretraining. It’s the foundation everything else is built on.

What it means

Pretraining is the initial training phase where a large language model learns from a vast dataset of text, typically books, websites, academic papers, and code, often totaling trillions of words. The model’s task during pretraining is simple in theory: predict the next token given everything that came before. It does this billions of times, adjusting its weights with each attempt, until it gets very good at prediction.

That prediction task requires genuine language understanding. A model that reliably predicts what comes next has learned grammar, facts, reasoning patterns, and world knowledge as a byproduct. Pretraining a frontier model takes months and costs tens to hundreds of millions of dollars, which is why only a handful of organizations can do it.

Why it matters

• Pretraining is where the model’s fundamental capabilities are set. Fine-tuning and RLHF can shape behavior, but they can’t add knowledge that wasn’t present after pretraining. A weak pretrained model produces a weak final model regardless of what comes after.

• The data used in pretraining determines what the model knows, what biases it absorbs, and what languages and domains it handles well or poorly. Decisions about what to include, filter, or weight in the dataset have downstream effects on every product built on top of that model.

• Pretraining is where most of the compute in AI gets spent, and it’s becoming more expensive, not less. Labs are training on larger datasets with larger models, pushing costs higher each generation. Andrej Karpathy joined Anthropic in 2026 to work on using Claude to improve Claude’s own pretraining, one of the clearest signals of how central this phase remains.

Simple example

Think of pretraining as the years a person spends reading before they ever write anything professionally. A person who has read widely across science, history, fiction, and technical manuals will write better first drafts than someone who hasn’t, regardless of the specific writing training they receive later.

Pretraining is those years of reading, compressed into months of compute, done at a scale no human could match.

It’s not AI related but...

My wife and I are launching a new service called PostARTing — an art subscription that mails original, handmade art postcards to your door. Real artists, real mail, no prints, and no AI art.

We opened a pilot to gauge interest and kick the tires and, well, it filled up in under half an hour.

So now we’re opening a limited Charter Club before we officially launch. Founding pricing locked in forever, plus a free card to send a friend. And there are just a few dozen spots left.

If getting real art in the mail sounds like your kind of thing, claim a spot before they’re all gone.

CLAIM YOUR SPOT

The AI industry produces thousands of models. Most are fine-tuned versions of existing ones, smaller distillations, or specialized tools built for narrow tasks. A small number are something different: the most capable models in existence at any given moment, trained at the edge of what current hardware and techniques allow.

These are frontier models. They set the benchmarks every other model is measured against, and attract the most attention, investment, and concern.

What it means

A frontier model is an AI model that represents the current state of the art in general capability, sitting at or near the leading edge of what the field can build. The term is deliberately relative: what counts as a frontier model changes as the field advances. GPT-4 was a frontier model in 2023. By 2025, dozens of models had matched or exceeded it on most benchmarks, and the frontier had moved.

Today’s frontier models include Claude Opus, GPT-5, and Google’s Gemini Ultra: more data, more compute, more sophisticated techniques than anything before. They’re almost always the most expensive to train, most capable across tasks, and most resource-intensive to run.

Why it matters

• Frontier models set the capability ceiling for the entire industry. Every fine-tuned model, every specialized tool, every API-based product is built on top of a base model, and that base model’s capabilities determine what’s possible downstream. When frontier model capability jumps, the whole ecosystem jumps with it.

• They’re also where the safety stakes are highest. A model operating at the frontier can do things previous models couldn’t, including things researchers haven’t fully anticipated. Most AI safety research focuses on frontier models because that’s where the risks are most acute and least understood. Labs that build frontier models operate under a different level of scrutiny than those building smaller or more specialized systems.

• Access to frontier models is increasingly a competitive moat. The compute required to train them is concentrated in a small number of organizations. The talent required to do it well is scarcer still. This is why the gap between frontier labs and everyone else tends to persist even as open weights models improve: the frontier keeps moving faster than it can be replicated.

Simple example

Think of frontier models the way you’d think of the fastest car in the world at any given year. Everyone else benchmarks against it. Engineers study it. Regulators watch it. The moment it’s surpassed, the new one becomes the reference point.

The frontier is always exactly one step ahead of the rest of the field, by definition.

AI models can’t read text the way you do. Under the hood, everything has to become math. Embeddings are the solution: a way of converting words, sentences, images, or almost any kind of data into lists of numbers that capture meaning, not just identity.

The result is that concepts with similar meanings end up near each other in mathematical space. “Dog” ends up close to “puppy.” “King” minus “man” plus “woman” gets you surprisingly close to “queen.” That’s embeddings working as designed.

What it means

An embedding is a numerical representation of a piece of data, typically a list of hundreds or thousands of decimal numbers called a vector, where the position and value of each number encodes something about what the data means. Models learn embeddings during training by processing enormous amounts of text and adjusting the numbers until similar concepts cluster together in the vector space.

Once trained, embeddings let you do things that raw text can’t: measure how similar two pieces of text are, find related concepts, power search engines that understand meaning rather than just matching keywords, and connect information across different data types. RAG systems, semantic search, and recommendation engines all depend on them.

Why it matters

• Embeddings are what allow AI to understand context rather than just pattern-match on words. A keyword search for “apple” returns everything with that word. A semantic search using embeddings returns results about fruit OR about the tech company based on what you actually seem to want. That difference is why modern AI search feels qualitatively different from older search.

• They’re the connective tissue between different AI systems. When you upload a document to an AI tool and it “remembers” what you uploaded, it’s almost certainly stored as embeddings in a vector database. When a chatbot retrieves relevant context before answering, it’s comparing embeddings. The architecture of most production AI applications depends on them.

• Embeddings leak information. Embeddings from sensitive text can sometimes be reverse-engineered to recover the original content. This is a real concern for any organization storing embeddings of private data.

Simple example

You run a customer support system with ten thousand past tickets. A new ticket comes in. Instead of searching for exact keyword matches, the system converts the new ticket into an embedding and finds the fifty closest embeddings in the database.

Those are the most semantically similar past tickets, regardless of whether they share a single word. The suggested responses are relevant because they come from actually similar situations, not just similar phrasing.

A model flags a loan application as high risk. A hiring tool ranks a candidate low. A medical system recommends against a treatment. In each case, the AI has an answer. What it often doesn’t have is an explanation. XAI is the field built to fix that: not just making AI accurate, but making it accountable enough that a human can understand, audit, and if necessary challenge what it did.

What it means

XAI (Explainable AI) refers to methods and techniques that make AI decisions understandable to humans. Where interpretability focuses on understanding a model’s internal mechanics, XAI is more applied: it’s about producing explanations that work for a specific audience, whether that’s a data scientist, a regulator, or an end user who just got denied a mortgage.

Common XAI techniques include LIME (Local Interpretable Model-agnostic Explanations), which approximates what a model did around a specific decision, and SHAP (SHapley Additive exPlanations), which assigns each input feature a contribution score showing how much it pushed the outcome in either direction.

Why it matters

• Regulation is driving adoption fast. The EU AI Act requires “meaningful explanations” for automated decisions affecting people in high-risk categories: credit, hiring, healthcare, law enforcement. In the US, financial regulators have required adverse action notices for credit decisions for decades. AI doesn’t change that requirement. It makes it harder to meet.

• Explanations catch errors that accuracy scores miss. A model can be 92% accurate overall while being systematically wrong for a specific demographic. Without explanations, that pattern stays hidden. With them, auditors can spot it. Several high-profile AI bias cases were only discovered because someone dug into the why, not just the what.

• XAI is now a product differentiator. Enterprise AI buyers, especially in regulated industries, increasingly ask vendors to demonstrate explainability before signing contracts. A model that can’t explain itself is a liability. One that can is easier to deploy, easier to audit, and easier to defend when something goes wrong.

Simple example

Two candidates apply for the same job. The AI ranks one higher. The lower-ranked candidate asks why. Without XAI, the answer is a score. With SHAP applied, the answer is specific: years of relevant experience contributed positively, gaps in employment history contributed negatively, and location had no effect. That’s a defensible explanation. It’s also something a human reviewer can check for bias. The score alone isn’t.

When an AI model gives you a wrong answer, or a biased one, or a dangerous one, the natural question is: what happened inside the model that produced that output? Right now, we largely can’t answer that. The model takes in text, runs it through billions of mathematical operations, and produces a response.

What it’s actually “thinking” along the way is opaque, even to the people who built it. Interpretability is the field trying to change that.

What it means

Interpretability (also called mechanistic interpretability) is the study of how AI models work on the inside: which parts of the model activate for which concepts, how information flows through the system, and what internal representations the model builds to produce its outputs. Anthropic has made interpretability a core research priority, publishing work on identifying specific “features” inside Claude that correspond to concepts like emotions, intentions, and abstract ideas.

In 2024, their researchers mapped features inside a small language model and found that individual neurons often represent multiple unrelated concepts simultaneously, a phenomenon called superposition that makes models significantly harder to understand than previously assumed.

Why it matters

• Safety depends on it. If we can’t see inside a model, we can’t verify that it’s actually doing what we think it’s doing. A model might produce correct outputs for the right reasons, or correct outputs for the wrong ones. Without interpretability tools, there’s no reliable way to tell the difference before something goes wrong at scale.

• Interpretability research has already produced useful findings. Anthropic’s work identified a feature inside Claude associated with the “Assistant” token that, when examined, showed signs of what researchers described as a kind of suppressed unease. Whether that reflects something real about the model’s internal state is an open question, but the fact that we can now ask it is new.

• Regulators are starting to require it. The EU AI Act includes provisions around explainability for high-risk AI systems: the ability to account for why a model made a particular decision. Interpretability is the technical foundation that makes those requirements possible to meet. Right now, most deployed systems can’t fully meet that bar.

Simple example

A radiologist using an AI diagnostic tool gets a flagged scan. The tool says cancer with 94% confidence. The radiologist wants to know what the model saw. Without interpretability, the answer is essentially: we don’t know, trust the number. With interpretability tools, you could in principle see exactly which pixels or patterns drove the prediction. Medicine, law, and finance all have this problem. The number isn’t enough. You need the reasoning.

Most tech CEOs sell optimism. Dario Amodei sells something more complicated: a bet that if powerful AI is coming regardless, it’s better to have safety-focused labs at the frontier than to cede that ground to teams less focused on the risks.

He co-founded Anthropic on that thesis in 2021, watched it grow to a near-trillion-dollar valuation, and hasn’t stopped warning publicly about what could go wrong.

Who they are

Amodei is the CEO and co-founder of Anthropic, which he started in 2021 alongside his sister Daniela Amodei and five other former OpenAI colleagues after leaving his role as OpenAI’s VP of Research. He holds a PhD in computational neuroscience from Princeton, and under his leadership Anthropic has grown from a small safety-focused research lab to one of the most valuable private companies in the world, valued at roughly $380 billion as of early 2026 and filing confidentially for an IPO in June 2026. Amazon has committed $25 billion to the company.

Amodei sits at the center of nearly every major debate in AI: safety, military use, and who should control the technology.

Why they matter

• Amodei is the most prominent voice arguing simultaneously that AI will be transformative and that it needs strict guardrails. At Davos in January 2026, he said he believes AI systems could surpass human capabilities “at almost everything” within a few years. He said it not as a sales pitch but as a warning that the window to get safety right is closing.

• He’s willing to take costly positions on principle. In early 2026, Anthropic refused to let the US Defense Department deploy Claude for certain military applications, resulting in the company being designated a “supply chain risk” by the Pentagon. OpenAI stepped in to fill the contract. Amodei publicly called the concentration of AI power in any small group, including his own, something he was “deeply uncomfortable” with.

• His 2023 essay “Machines of Loving Grace” outlined a vision of AI compressing decades of scientific progress into years, potentially defeating diseases and lifting billions out of poverty. It’s the most detailed optimistic case for AI written by anyone actually building it.

What they’ve said or done

At Davos in January 2026, Amodei told the Wall Street Journal:

“I think it’s plausible it could be longer than 2027. I don’t think it will be a whole bunch longer than that when AI systems are better than humans at almost everything.”

He wasn’t predicting a utopia. He was describing a deadline.

Most software is predictable. Run the same calculation twice and you get the same result. AI language models don’t work that way. Ask the same question in the same session and you’ll often get a somewhat different response each time: different wording, different examples, sometimes different conclusions.

This property has a name: non-determinism. It’s one of the most misunderstood things about how AI works, and it matters the moment you start building with it.

What it means

Non-determinism in AI means that a model doesn’t always produce the same output for the same input. The reason is a setting called temperature, which controls how much randomness the model injects when choosing its next word. At temperature zero, the model always picks the most statistically likely next word, producing consistent, predictable output. At higher temperatures, it samples from a range of likely options, introducing variation. Most consumer AI products run at a temperature between 0.7 and 1.0, which is why responses feel natural and varied rather than robotic. You can’t guarantee identical output twice, even with an identical prompt.

Why it matters

• Non-determinism is a feature in creative work and a liability in production systems. It’s what makes AI writing feel human rather than templated. But if you’re building a system that needs to pass the same test every time, return consistent structured data, or behave predictably in a regulated environment, non-determinism is a real engineering problem you have to design around.

• It makes testing AI systems genuinely hard. Traditional software testing assumes the same input produces the same output. With non-deterministic models, a test that passes today might fail tomorrow with the same inputs. Teams building production AI systems often run hundreds of sample outputs to characterize behavior statistically rather than test for exact matches.

• Temperature isn’t the only source of non-determinism. Model updates, infrastructure changes, and even parallel processing at the hardware level can all introduce variation. Setting temperature to zero reduces randomness but doesn’t eliminate it entirely. True determinism in a large language model is harder to guarantee than it sounds.

Simple example

A law firm builds an AI tool to draft standard contract clauses. In testing, the output looks great. In production, they notice the same clause prompt generates slightly different language each time: same meaning, different phrasing. For a creative brief that’s fine.

For a legal document where consistency matters, it’s a problem. They switch to temperature zero and add a review step for any output that will go into a signed document. Non-determinism managed, not eliminated.

When an AI agent takes on a complex task, it faces a real constraint: its context window fills up fast. The longer it works, the more history it carries, and the more likely it is to lose track of earlier decisions or slow down. Subagents solve this by breaking the work into pieces. Instead of one agent doing everything in one long session, a lead agent delegates chunks of work to specialized subagents, each operating in its own clean context. They report back. The lead agent synthesizes the results.

What it means

A subagent is an AI agent instance spawned by a parent agent to handle a specific, isolated piece of a larger task. Each subagent gets its own fresh context window, works independently, and returns its output when done. The parent agent, sometimes called the orchestrator, coordinates the subagents: deciding what to delegate, in what order, and how to combine the results.

In Claude Code, subagents are a first-class feature. The orchestrator spins up subagents for discrete steps like research, planning, and implementation, running them sequentially or in parallel. Each subagent starts clean, with only the context it needs.

Why it matters

• Subagents are the most effective architectural pattern for avoiding context overload in long agentic sessions. When each phase of a task runs in its own context window, none of them accumulate the full weight of everything that came before. The work stays focused, and the model stays reliable across the full session.

• Parallel subagents dramatically cut the time complex tasks take. A research task that requires gathering information from ten sources can spawn ten subagents simultaneously rather than running each source sequentially. The parent agent waits for all ten to finish, then synthesizes. Tasks that would take an hour can take minutes.

• Subagents also contain failure. If one subagent goes off the rails on a subtask, it doesn’t corrupt the parent session or the other subagents. The orchestrator can catch a bad result, discard it, and retry that piece without restarting the whole job.

Simple example

You ask an AI agent to research three competing products, compare their pricing, and draft a recommendation. A single-agent approach reads source one, then source two, then source three, carrying all of it in one growing context.

A subagent approach spawns three research subagents in parallel, one per product, each starting fresh. They finish, hand their summaries to the orchestrator, and the orchestrator writes the comparison from clean, structured inputs. Same output. Much less context drag.