Sixty years of telling computers exactly what to do. Then, very recently, something else. This page is everything from the session: the three moves, what the machine can and cannot do right now, twenty things worth trying, and what to read so you stay current without drowning.
They work on anything: a job application, a dissertation, a pitch, a decision you are stuck on. Any AI will do. Talk to it rather than type, you will tell it more.
Numbers. Give it a spreadsheet or ask it to write the calculation. It is reliable when it writes the code. Never trust arithmetic it did in its head.
How it has always been. Software does exactly what it is told. Same input, same output. You are the intelligence, and you are the bridge between every system that does not talk to the next.
What is different. A model is not a calculator. Ask it the same thing twice and you get two answers. You cannot remove that spread, and you would not want to. You narrow it, with your context, the right tools, a fence around what it may do, and your own review of what comes back. All of that around the model is what people mean by the harness. Anthropic's shop experiment showed it plainly: the same model that claimed to be human and sold at a loss ran three shops at a profit once it had an inventory that showed cost, a rule to check prices, and a boss.
How good it is. The best systems now complete roughly a working day of human effort at a fifty percent success rate, a few hours at eighty. Doubling every three months. And it is jagged: better than the best humans at competition maths, under one percent on a game it has never seen, thirty points behind you at using a computer, unable to read a clock.
Watch it work. Delegation, shown: proposals in, a memo out, one conclusion traced to its source. Then the creative side: an interface drawn while you use it, a goal replayed from an angle no camera filmed, your own voice heard in another language.
This industry. Performance analysis is mature. Officiating is deployed. Content is moving fast. Commercial and operations have barely started, and that is where you will work. The league in this city has decided everyone has to learn this.
You. Several agents inside one fence, sharing the tools, with a budget, and one person above it setting the goal. The skill is not technical. It is knowing what to ask for, and recognising when the answer is wrong.
of people using generative AI within three years of its arrival. Faster than the PC, faster than the internet.
of human work completed alone at a 50% success rate. A few hours at 80%.
cheaper for the same quality of answer in about eighteen months, and still falling.
of the time Liverpool's own analysts preferred the machine's corner routines to their existing ones.
higher success rate after six months of ordinary use. It holds after controlling for the task. Learning by doing is real.
agent workdays for every human workday inside OpenAI's research organisation, since June 2026. Their caveat: machine time, not results.
Intelligence is jagged. The same model on the same day is brilliant at some things and worse than a child at others. Check the things it is bad at. Stop checking the things it is reliably good at.
Ranked for how much they teach, not for benchmark scores. Each one says what you would see, where it earns its place, and the honest limit. Checked on 13 September 2026; product versions and access move quickly, so confirm before you rely on one.
An interface generated while you use it. Solaris generates interactive interface frames in response to clicks, drags and prompts. Runway shows shopping, room manipulation and interactive instructional scenes. A language model directs how the generated interface evolves.
Where it earns its place: For marketers: personalised product exploration. For builders: a new way of representing interaction, and training environments for agents that use computers.
The honest limit: Use the provider recording. Stable text, reliable facts, coherence over long sessions and accessibility remain open challenges; a convincing screen does not establish a working commerce backend.
A world that changes as you explore it. Runway demonstrates interactive worlds with generated video and audio, camera movement, character actions and scene changes. Its stated output is continuous 720p video at 24 fps, with 48 kHz audio.
Where it earns its place: Interactive entertainment, experiential campaigns, training environments and simulation research. Students can compare a generated world with a programmed game.
The honest limit: The clips are provider demonstrations. No preset session length is not a promise of indefinitely accurate memory or physics; difficult motion and consistency over time remain limitations.
Have a video conversation with an AI character. The newest Phoenix release generates coordinated face, head, shoulder and torso movement while an AI character speaks and listens. Tavus says existing stock Faces have been upgraded and provides a public model demo.
Where it earns its place: Sales practice, interactive tutoring, onboarding and concierge experiences. The technical lesson is how perception, conversation and rendering combine into one experience.
The honest limit: Realistic social behaviour does not demonstrate human understanding. Treat Tavus performance and engagement numbers as vendor claims; do not extrapolate them to your audience.
An AI voice that changes tone with the conversation. Expressive Mode combines Eleven v3 Conversational speech generation with better turn taking. The announcement includes an unedited example responding to a frustrated passenger after a flight cancellation.
Where it earns its place: Brand voice, multilingual support and realistic conversation practice. Students can hear how timing, interruptions and delivery change the quality of an interaction.
The honest limit: This is an older release retained for its teaching value. Speech that sounds emotional does not establish empathy or the accuracy of the underlying answer.
Build a playable world through iterative instructions. OpenAI developer Thomas Ricouard documents building Void Explorer in Codex: flying between stars, descending to planets and walking on their surfaces. The account shows art direction, implementation decisions and testing.
Where it earns its place: Students can move from an idea to software they can inspect and modify. Business audiences can see interactive prototypes, calculators and demos as practical outputs.
The honest limit: This was an iterative build with human direction and validation. Do not describe it as a finished game produced reliably by a single prompt.
Delegate work across documents, tools and a browser. Cowork supports tasks across connected tools and files, with sessions available on desktop, web and mobile. Cloud work can continue when a laptop closes; local file, browser and computer access depends on the desktop connection.
Where it earns its place: A clear business example of delegation; a useful student lesson in specifying outputs, evidence and review criteria rather than asking for a generic answer.
The honest limit: Features differ by plan, surface and rollout. For a live demonstration in the browser or from local files, keep the desktop app online and rehearse the exact workflow.
Change one frame and carry the edit through a video. Aleph 2.0 applies an edited frame across video, including relevant shots in sequences of several shots. Runway documents clips up to 30 seconds at 1080p and examples such as product colour, clothing and background changes.
Where it earns its place: One campaign asset can become several creative treatments. For students, this makes temporal consistency and precise editing visible without a technical lecture.
The honest limit: Inspect unchanged regions for unwanted changes. The promise of preserving the rest is a product objective, not a guarantee on every generated frame.
Revisit a scene from a camera angle you never filmed. Atlas combines text, images, video and 3D in a spatial model. World Labs demonstrates generation with camera control, reconstruction from sparse views, and reframing video from new viewpoints.
Where it earns its place: Virtual location scouting, visual effects, property experiences and simulation. The teaching distinction is spatially controlled reconstruction versus ordinary image animation.
The honest limit: Atlas is entering early access with selected partners and will power future products. Marble is the public alternative, but using Marble is not the same as using Atlas.
Create a song, then direct a specific musical change. Suno introduced v6, v6-wild and v6-mini. The family supports music creation from multiple input types and edits in plain language to parts of a song, including changing a lyric or chorus.
Where it earns its place: Brand sound exploration, rough campaign soundtracks and student creative projects. The memorable moment is the editable musical decision, beyond simply hearing an AI song.
The honest limit: v6 and v6-wild require paid access; mini has a different access tier. Use original material and check the applicable product terms before commercial reuse.
Direct a video with explicit visual endpoints. Google documents scene extension, first and last frame control, quick previews at low resolution and 4K upscaling in Omni 1.1 Flash. These controls make iterative video production easier to demonstrate.
Where it earns its place: Marketers get a concrete story about creative control. Builders can discuss the workflow from cheap previews to more expensive final output.
The honest limit: 4K upscaling is different from native 4K generation. Check the specific controls and model available in the product or API account you plan to demonstrate.
A speaking assistant that can reason and call tools. GPT-Realtime-2.1 supports spoken conversation, speech in and speech out, configurable reasoning effort and tool use. The model page reports improvements in alphanumeric recognition, silence, noise handling and interruptions.
Where it earns its place: Voice interfaces for service workflows and interactive applications. Students can distinguish a component that turns text into speech from a model that reasons and uses tools.
The honest limit: This needs a configured application or playground. API model availability does not establish which model powers consumer ChatGPT Voice; more reasoning may add latency.
Direct a voice in plain language, like an actor. Realtime TTS-2 conditions delivery on conversation context and supports voice direction in plain language and multilingual speech. Inworld provides voice examples side by side and a live pipeline demo.
Where it earns its place: Branded voices, interactive characters and customer conversations. This is an immediately audible example of control over delivery, not just text generation.
The honest limit: Inworld pages use different language counts and latency definitions. Avoid one headline speed claim: synthesis time excludes other parts of the conversation pipeline.
Speak one language and hear another in your vocal style. Google describes continuous speech translation across more than 70 languages, preserving aspects of intonation, pacing and pitch. The launch covers Translate mobile apps, developer access and a separate Meet preview.
Where it earns its place: International events, multilingual collaboration and language learning. It connects an obvious daily problem to a capability audiences can immediately understand.
The honest limit: It remains a few seconds behind the speaker and can mistranslate. Availability differs across Translate, Meet and the API; verify the chosen language pair and account.
Art direct an image, step by step, with precise changes. Sunburst generates and edits images using text and image inputs, with OpenAI positioning it for workflows where editing precision matters. It is directly selectable through the image API and image generation tool.
Where it earns its place: Campaign concept development and visual prototyping. Students see why reference images, constraints and checking details matter as much as the initial prompt.
The honest limit: A polished image can still change text or object details. Do not infer that every consumer image interface exposes this exact model or all API controls.
Turn governed business data into a working application. Replit now offers generally available Databricks integration with native Lakebase support. The workflow combines application creation, governed business data, preview environments and automated database provisioning.
Where it earns its place: Business teams can envisage custom internal tools; students can see how an application connects an interface, stored data and access rules.
The honest limit: A demo using a local sample dataset does not demonstrate the Databricks integration. The enterprise workflow requires configured accounts and its own approvals.
Turn a source pack into an explainer and an analysis. Google renamed NotebookLM to Gemini Notebook and began adding a cloud computer that runs code on your sources. Its established output formats include audio and video overviews, slide decks and infographics.
Where it earns its place: Research briefings, onboarding and study aids. This is a useful bridge between the business and education halves of the audience.
The honest limit: Citations improve inspectability, not guaranteed accuracy. The rename is newer than many tutorials; the cloud computer and other advanced features have separate plan and rollout conditions.
Give a persistent agent a standing responsibility. Grok Bot is organised around persistent roles, computers, memory and routines. The design article describes specialist Bots working together and tasks that run on a schedule or when an event happens.
Where it earns its place: Recurring research and operating workflows. For students, it makes the distinction between a model, a single task and a persistent agent concrete.
The honest limit: The source is an account of the product design, not an independent reliability study. Use its walkthrough unless your account and the exact routine have been rehearsed.
An agent can pay for a service it needs. MPP lets agents and services coordinate payments programmatically. Stripe describes examples including paying for a browser session, sending physical mail and ordering a sandwich.
Where it earns its place: A new customer type for businesses: software acting on an authorised user's behalf. Builders can see why payments and permissions belong in the application design.
The honest limit: The protocol enables transactions; it does not itself establish that a purchase is useful or authorised. Keep a live classroom example in a prepared test environment.
AI coordinates a robot's whole body to complete work. Google demonstrates control of the whole body and dexterous manipulation across different robot embodiments. Figure's Helix 02 provides a separate concrete example: a four minute autonomous dishwasher task, according to Figure.
Where it earns its place: A tangible discussion of physical work, logistics and embodied AI. It gives the audience a break from screens while connecting to familiar household tasks.
The honest limit: These are company demonstrations under particular conditions. A successful recorded task does not establish general home reliability or consumer availability.
An agent searches for better algorithms and tests them. AlphaEvolve combines proposed code changes with automated evaluation to search for better algorithms. Google reports applications in its infrastructure and broader scientific work, and July availability for Cloud customers.
Where it earns its place: Operations optimisation and research. Students see why a measurable evaluator can make an agent more useful than a fluent answer alone.
The honest limit: The objective and evaluator constrain what can be improved. Do not turn a specific verified optimisation into a claim that AI can solve arbitrary scientific problems.
Three you can check. DeepMind and Liverpool's TacticAI for corners, Nature Communications 2024. Semi automated offside, live since Qatar 2022. Automated highlights from WSC Sports and others in most major leagues.
What to unlearn. Same input, same output: it will not be. The first answer is the answer: it is a first draft of a spread. Asking questions: give it briefs instead. A question gets an answer. A brief gets work.
What to learn. Framing a vague need into a precise, checkable task. Checking fast, and knowing what to check. Knowing what good looks like in your field, and keeping some work for yourself so your judgement stays sharp.
If you subscribe to five: The Rundown for the daily scan, ThursdAI for new demonstrations, One Useful Thing for work and learning, Latent.Space for technical depth, and Simon Willison for hands on testing.
Daily. Quick product news, use cases and practical guides. A good first scan for a mixed business and student audience. Follow the original announcement before quoting a capability or release date.
Weekly show + newsletter. Live discussion of model launches, often with tests, timestamps and links. Its monthly release index is particularly useful for finding candidates. The episode or index date can be later than the actual announcement.
Essays. Explanations, informed by research, of how AI changes tasks, organisations and learning. Use it for the argument around a demo. It is interpretation and experimentation, not an exhaustive release feed.
Newsletter + technical podcast. Conversations with builders about agents, models, infrastructure and AI for science. Useful preparation for questions from technical students. Longer and more technical than a headline digest; select relevant episodes.
Frequent posts + subscription options. Concrete experiments, code and observations about model behaviour and agent engineering. An excellent counterweight to a polished launch video. A blog and subscription feed rather than a conventional daily news digest.
Weekly. Accessible explanations of research and industry developments for practitioners, executives and learners. The direct site blocked retrieval; its official indexed about page was available. Check the live archive for the newest issue.
Newsletter + tutorials. Tools, practical experiments and a founder perspective. Useful for finding an exercise that students or business teams can repeat. Its current Substack publication differs from older descriptions of a simple daily roundup.
Frequent newsletter. Readable product updates, tool discovery and practical learning material for people new to the subject. Use it as an alternative to another general digest to avoid repeated headlines.
Daily. Short news, tools and tutorials for work; especially relevant to marketers looking for an immediate experiment. Separate editorial examples from sponsored products and promotional claims.
Weekdays. Short summaries of model releases, research and developer tools, with links back to sources. Denser than the general newsletters; useful alongside one accessible source.
Weekly. Detailed research interpretation and discussion of the wider consequences of capability progress. Clark is a cofounder of Anthropic; read it as an informed perspective with a disclosed industry affiliation.
Weekdays. Aggregates discussion from AI X/Twitter, Discord and Reddit. Particularly useful for finding the experiment behind a circulating headline. Treat community summaries as leads and verify important facts at the primary source.
You bring the judgement.
It brings the hours.
This session, its diagrams and this page were built by directing AI, then checking every number against a source and throwing away what was wrong. That is the shape of the work now.
This page lives at 1706.ai/wfs. Point a phone at the code, or send the link.