Taste: The main advantage in AI – Tejas Kumar (Talk Outline)

A Craft 2026 talk by Tejas Kumar (pronounced “like contagious”) — a developer advocate for AI at IBM who studied human-computer interaction (HCI) at Stanford. His framing “why”: the way we interact with computers (HCI) is changing rapidly. The talk is deliberately academic and citation-backed (peer-reviewed research, not just opinion). Structure: where taste comes from → how it goes wrong → a working definition with worked examples → why it matters for business → three protocols to train it → Pixar’s Brain Trust → how taste shows up in engineering (performance/accessibility/time-to-value) → live demos (streaming, MCP, MCP apps) → close + Q&A. He opens by thanking sponsor Lufthansa (“cool to see Germany represent”) and jokes about the frozen clock (“it’s going to be a long talk”).

1. Why Taste, and Why Now

1.1 Anyone can make “some AI approximation of anything”

• Nowadays most people can make most things — “anybody can make anything.”
• His friend Titus corrected him the night before: not anybody can make anything, but anybody can make some AI approximation of what they want.
• With tools like Claude Code, Cursor, Codex, we have so much power that we’d better build things that are good and have taste — that care about the people on the other side of the screen — rather than slop.

1.2 “Slop” was the 2025 word of the year

• Slop was the word of the year for 2025.
• Slop is what they serve in prison cafeterias — food that “doesn’t taste like anything.”
• This talk is about the opposite of slop: taste.

1.3 The journey ahead

• Explore taste through science and engineering: where it comes from, how it goes wrong, and how to build/train your own taste and become a “tastemaker” (what gives Vogue the right to say an outfit is good or bad?).

2. Where Taste Comes From

2.1 Nobody is born with taste

• (Audience mostly doesn’t believe we’re born with taste; one person did — “I’m about to prove you wrong.“)
• Nobody is born with taste — it’s developed over time.

2.2 We are born with the precursor: pattern-matching

• Taste is a refining of neural circuitry we’re all born with.
• Taste begins in infancy — we’re born with pattern-matching mechanisms.

2.3 The newborn example

• A newborn gets smacked, oxygen enters the lungs, they cry, a caregiver gives care.
• Babies start pattern-matching: “if I cry, they feed me”; “if I’m sleepy, they make me sleep” — if this, then that.
• Same as us: “I write a prompt, I get software.”

2.4 The flaw in the system

• We assume “if a pattern is easy, it must be right” — if I can make sloppy garbage software with no effort, it must be good.
• We assume the easy choice is the right choice, and sometimes it isn’t.

2.5 How pattern-matching goes wrong

• Broken pattern-matching gives rise to prejudice, racism, sexism.
• Personal example: because of how he looks, people think he works in tech support, and have literally confused him with Banjo (fellow speaker Banjo Obayomi).
• We think “this feels about right / the easy choice must be right,” leading to systematic inequality and prejudice.

2.6 Taste = pattern-matching corrected by science

• Taste is pattern-matching, but corrected according to science — “when we get pattern-matching right.”
• The claim of the talk: taste starts as pattern-matching, goes through a refining process, and refined/redeemed/restored pattern-matching becomes taste.

3. A Working Definition of Taste

3.1 The definition

• Taste is a trainable skill where we are able to tacitly judge quality based on external standards.

3.2 “Trainable skill”

• You’re not born with it — you acquire, refine, and grow it; it keeps evolving.

3.3 “Tacitly”

• Tacitly = a gut feeling you can’t put into words.
• The French call it je ne sais quoi — “I don’t know what it is, it just feels right.”

3.4 “External standards” (the big one)

• With pattern-matching we make tacit judgments — but often based on internal bias (broken pattern-matching).
• Taste is when the same gut-feeling judgments are based on external standards — facts, science, things we know — not “I think all Indians are tech support.”

4. Worked Examples of Taste

4.1 Counterfeit money experts

• The clearest example: money. People whose job is to identify counterfeit money — a task “we still don’t trust AI with fully.”
• They make tacit judgments (“that’s fake, that’s real”).
• The naive assumption: they study all the counterfeits/scams. The truth: they study real money — they “live, breathe, smell, taste” genuine government-issued cash.
• Because they know the true thing so well, they instantly know the fake — a tacit judgment based on the external standard of real money.

4.2 The creative director and centering

• His first job was at a creative/design agency with a talented creative director, Stephanie.
• She gave him a design; it required centering something. He wrote align-items: center (a flexbox) and said “it’s done.”
• She said ”that’s not centered” — he argued “CSS doesn’t lie / computers are logical.”
• Investigating, he found it wasn’t the code — it was the font, which had extra height under each letter pushing things slightly higher.
• She was right because she tacitly knew — from immersion in “a billion vertically aligned things” — the external standard.

4.3 The off note (relative pitch)

• Tejas is a musician (guitar, singing; former singing coach).
• At a concert everyone’s in harmony, then the singer hits one off note (“Happy birthday… oh, Jesus”) and everyone cringes.
• We tacitly know the note doesn’t work because of the external standard of relative pitch — in the key of B you can use G# minor, F#, E, but an A major and “you’re cooked”; A♭ minor works.

5. Why Taste Matters

5.1 Personal

• He personally wants good taste and to do things people like.

5.2 Business: perceived value → profit

• If you build a product, taste makes it look higher value than it actually is → profit, and it influences purchase intent.

5.3 The Apple example

• Apple shows you their taste: “our new iPhone is made from the finest titanium,” zooming in on machining and glass.
• This raises perceived value: it may cost ~$500 to produce, but the beautiful video makes you willing to pay far more.
• Taste raises how much we rate things → how much we’ll spend — key for business, and personally.

6. How to Train Taste — Protocol 1: Immersion + Cues + Corrective Feedback

6.1 Taste is trainable — this is the payoff

• The point he most wants to land: taste is trainable — you can leave today and improve it, so the things you create are tasteful and considerate.

6.2 Full immersion

• You need full immersion (like Stephanie with aligned things, or the money experts with genuine cash).

6.3 Immersion alone is an echo chamber

• Immersion alone isn’t enough — you could immerse in a social-media echo chamber where cronies refeed your own bias and you become “delusional” (jab at “any world leader”).

6.4 Add valid cues

• You need valid cues — the ability to look at things and judge them.
• But cues alone make you ”God on a throne” declaring good/bad — not enough.

6.5 Add corrective feedback

• The final element is corrective feedback — an authority (ideally facts) saying “this judgment was off, and here’s the data.”
• Source: Dr. Kahneman and Klein, 2009 paper “Conditions for Intuitive Experience: A Failure to Disagree.”
• Formula: immersion + valid cues + corrective feedback.

6.6 The missing ingredient: humility

• Not in the science, added “for free”: humility — the ability to look inside and say “maybe I don’t have all the answers; I’m open to adjust.”

7. Protocol 2: Externalize a Standard

7.1 Dieter Rams and Braun

• Dieter Rams was the main designer at Braun (razors, accessories, peripherals), known for design.
• He wrote “Ten Principles of Good Design”, which still affects the world today.

7.2 Write down your standard and consult it

• Create your own standard from the taste you’ve built, write it down externally (“this is what my values/taste are”), and consult it when you make something.

7.3 Jony Ive and Apple

• Jony Ive (Apple’s chief of industrial design) was a student of Dieter Rams’ book and did exactly this with the iPhone, MacBook, and more.
• He took Braun’s razor design language and applied it to today’s iPhone.

8. Protocol 3: Cross-Pollinate

8.1 Steve Jobs’ typography class

• Whatever you think of him personally, Steve Jobs had incredible taste — the reason Apple is a trillion-dollar company.
• He dropped out of college but sat in one typography class, captivated by fonts (“the tail of the Y under the baseline”).
• He cross-pollinated that taste into the MacBook, iPhone, and Watch.

8.2 Apply lessons across domains

• You won’t get everything from a tech conference, but you can take lessons and apply them elsewhere — Tejas applies a ”pragmatic engineering mindset” to his sports club.
• Cross-pollination is a great way to cultivate taste.

9. Taste Perceives Differently (Not Just Judges Better)

9.1 The one-sentence summary

• All protocols summarize to: taste doesn’t just judge better, it perceives differently entirely — “you’re not even playing the same game.”

9.2 The call-to-action button experiment

• Mentoring junior engineers (front-end/React), he’d ask them to comment on the “Buy Now” button.
• Juniors comment on the button itself: “I like it, brand color, nice and big, good typography.”
• Seniors/staff say almost nothing about the button — instead: “on a slower network this button doesn’t show up at all,” or “on some phones it’s below the scroll fold.”
• Trained taste judges the environment around the button, not just the button.

9.3 The harness parallel

• This is what an AI harness does — it controls the entire environment around the agent/model (he has a separate YouTube talk on agent harnesses).
• Apple captured this with the old “Think Different” campaign — i.e., perceive differently = have taste.

10. Pixar’s Brain Trust

10.1 Creativity, Inc.

• Pixar (Toy Story, Finding Nemo) wrote “Creativity, Inc.”, a memoir documenting their creative process.

10.2 The Brain Trust practice

• They’d gather writers and filmmakers in a big circle to comment on each other’s work.
• Filmmakers were invited to bring their “ugly babies” — their unfinished/beta movies not yet refined.

10.3 Rule 1 — diagnose, don’t prescribe

• Feedback was not prescriptive (“change this, do that”) — instead ”I don’t get it,” “this didn’t taste right, something’s weird.”
• Rule: no prescriptions, only diagnosis.

10.4 Rule 2 — no hierarchy

• No hierarchy — directors and executive producers were present but never pulled rank; it was “my opinion, good luck,” always up to the filmmaker what to do with it.

10.5 Why it works

• It has all the ingredients: immersion (looking at each other’s ugly babies), cues (others give feedback), and corrective feedback (filmmakers take it on) — the full loop.
• Pixar used this exact process on Finding Nemo (he still remembers Dory, the blue fish).

11. Taste in Engineering: The External Standards

11.1 Performance and accessibility

• Almost every engineering discipline expresses taste objectively through performance and accessibility on the web.

11.2 Performance primitives

• Input to Next Paint (INP).
• Largest Contentful Paint (LCP) — how long the largest portion of your site takes to load.
• Cumulative Layout Shift (CLS) — when something loads and pushes a button you’re about to click somewhere else.
• These are measured with numbers.

11.3 Accessibility standards

• WAI-ARIA (WAI = Web Accessibility Initiative) and WCAG (Web Content Accessibility Guidelines) — external standards for whether your thing is tasteful.

11.4 The proposed standard: time to value

• His informal proposed standard: time to done / time to value — how fast someone gets work done with the software.
• Examples: on Expedia, how long to buy a ticket or book a hotel; on Lufthansa, how long to choose a seat without clicking through unwanted insurance offers.

11.5 “Milliseconds make millions”

• Google’s article “Milliseconds Make Millions”: a 100-millisecond (0.1s) improvement raises conversion rates by significant orders of magnitude — freely available on web.dev.
• Taste is literally the difference between slop and value.

12. Demo 1 — Streaming (Time to Value)

12.1 The old way

• (Everyone in the audience has used an AI-generation app like ChatGPT.)
• The old way: send a request and wait seconds/minutes for the model to generate a full blob of text before it’s sent over the network.
• Live demo (“write a short poem about Budapest”) took 4.6 seconds — “I would have closed the tab.”

12.2 The new (streaming) way

• Same request, but streaming: text appears immediately — it actually took longer in total, yet felt better because you saw something on screen.

12.3 Time to value ≠ time to completion

• Time to value doesn’t mean objective time to completion — here the objective time is worse.
• People want to feel like their stuff matters and is working (long waits make people open dev tools thinking something’s broken — a failure mode).

12.4 The code difference

• Old code: get button → on click, get prompt + API key → fetch an AI service → await turning the response into JSON (the blocking step).
• New code: same setup, but don’t await JSON — the response body is a stream of binary bytes.
  • Pipe it through a TextDecoderStream (binary → ASCII/English characters).
  • Get the stream’s reader; in a while (true) loop, read — if done, break; otherwise concatenate the value (add a line on newlines).
  • Append only the delta (the changed characters) to out.textContent, not the full string.
• “Small difference, huge impact” — the name of the game with taste.

13. Demo 2 — MCP (“Bring the info to my house”)

13.1 The pain of the modern web

• He doesn’t enjoy using the web (living in Europe: cookie banners, “continue with Google,” constant decisions).
• Demoing Expedia / adidas.de: sites make you make decisions before giving any value — “the web trains us to say no.”
• The reason: they weren’t thinking of you when they built it — they were avoiding fines or serving “business nonsense.”

13.2 The Craft Conference site

• He loves the Craft Conference website (not obtrusive), but finds the agenda “quite a lot,” especially on a phone (a bottom element “gives me anxiety” — said respectfully).

13.3 The idea: information comes to your house

• The web makes you go to someone else’s house (“give me the website that lives here”).
• What if the information came to you — no one else’s carousel/cookie-banner decisions?

13.4 What MCP is

• Model Context Protocol (MCP) lets any AI app (ChatGPT, Claude, Cursor, Windsurf) connect to an external MCP server and receive context.
• The server provides tools (callable functions), context, and prompt templates.

13.5 Building an MCP server (~100 lines)

• Easy to make: create a new MCP server, register a bunch of tools, start the server, connect the transport — “just over 100 lines of code.”
• Each tool has a name, metadata, and a description (the description is how the LLM knows to call the tool), plus a function implementing the data retrieval.

13.6 Connecting it to Claude desktop

• Adds his server to Claude desktop via developer settings — in 2026 you edit a JSON file (“I hope it gets better”).
• Aside/jokes: Claude doesn’t quit after you click quit (“they spy on you… you have to quit it twice… if Anthropic is listening, fix that”).

13.7 The result — and what’s lost

• He asks ”when is Banjo’s talk?” → the server queries the Craft site and returns Banjo Obayomi’s talk: “Beyond Vibe Coding: Building the Harness…” — “that’s all I want,” no cookie banners.
• But something is lost: the brand. Craft has a logo, feel, and colors — “this has become less human”; we go to a filmmaker’s movie for their taste — “I’ve lost their taste.”

14. Demo 3 — MCP Apps (Bring the Taste Back)

14.1 MCP apps extend the standard with UI

• MCP apps let engineers extend their MCP server to include UI they create.

14.2 Registering an HTML resource

• Register a resource (like a big text file given to the LLM) — similar to a tool.
• Extra fields: a URI (an ID like schedule://widget, just a string to identify the resource) and a MIME type text/html; profile=...mcp-app telling ChatGPT “this is an MCP apps UI.”
• The content is a big blob of HTML (schedule-widget.html) — plain HTML he wrote himself, styled with Craft Conference branding.
• Then create the server and listen on a port.

14.3 Connecting to ChatGPT

• Runs npm run chatgpt, exposes his machine to the internet via a tunnel endpoint (“please don’t hack me”).
• ChatGPT supports MCP apps: go to chatgpt.com/apps → create your own, name it “craft conf,” paste the tunnel URL (no auth, though it supports OAuth), create it.
• ChatGPT does a handshake (“are you a real MCP server?”) and connects.

14.4 The tasteful result

• Naively asking “who is speaking” is “lame” (a deliberate trap — one attendee was paying attention).
• The right way: @-mention the MCP server (“who is speaking at…” / “when is Titus’s talk?”) to tell ChatGPT to use his server.
• Result rendered inside ChatGPT with the actual Craft Conference logo and branded UI — e.g., ”Titus Winters, senior principal scientist at Adobe,” with the yellow stage shown in yellow.
• ”Tasteful UI inside my house — I just prompt and get what I want, and I still maintain the taste of the author.”
• Quick contrast (alt-tab between the plain vs. branded “when is Banjo’s talk?”) shows the difference — bringing taste back into soulless AI.

15. Close

15.1 Recap

• Taste begins at birth as pattern-matching → can go wrong (unconscious bias, prejudice, racism, ego) → refined with external cues + corrective feedback → becomes quality taste.

15.2 IBM and open-rag

• ~375,000 people work at IBM; you can’t have taste across the whole org, but his small team tries to.
• Their open-source project open-rag helps enterprise customers operationalize their internal data (NDAs, access rights) — e.g., a customer can chat with last year’s team call (a video) and AI answers — done as tastefully as possible.

15.3 The commission

• “Go forth and create really beautiful and tasteful things.”

16. Q&A

16.1 Q1 — Valid cues + immersion (clarification)

• Valid cues + immersion is like money — you look at genuine and counterfeit articles and judge them, as the Pixar people do.
• But immersion + valid cues alone is not enough — you need corrective feedback from an authoritative source.

16.2 Q2 — Since taste comes from immersion, should we teach students by reading and critiquing good code and systems?

• Absolutely yes.
• He brought his friend Jonathan, who’d never been to a tech conference — “beautiful to watch his taste grow” (“this talk I got this, that talk I got that”).
• This is why Montessori schools work — you’re exposed to real cues.

16.3 Q3 — the withdrawn question

• Someone submitted a question and left; the MC said it was “too personal” — so they wrapped up (audience laughter).

People & References Cited

• Tejas Kumar — speaker; developer advocate for AI at IBM; HCI, Stanford; musician/former singing coach.
• People: Stephanie (his first creative director), Titus / Titus Winters (senior principal scientist at Adobe; corrected him on “anybody can make anything”), Banjo Obayomi (fellow speaker — “Beyond Vibe Coding”), Jonathan (his friend), Dieter Rams (Ten Principles of Good Design, Braun), Jony Ive (Apple), Steve Jobs (typography class), Daniel Kahneman & Gary Klein (2009 paper).
• Companies/products: Lufthansa (sponsor), IBM (open-rag), Apple (iPhone, MacBook, Watch, “Think Different”), Braun, Pixar (Toy Story, Finding Nemo), Vogue, Expedia, adidas.de, Google/web.dev, OpenAI/ChatGPT, Anthropic/Claude (Claude Code, Claude desktop), Cursor, Windsurf, Codex.
• Papers/books/articles: Kahneman & Klein, “Conditions for Intuitive Experience: A Failure to Disagree” (2009); Dieter Rams, Ten Principles of Good Design; Pixar, Creativity, Inc.; Google, “Milliseconds Make Millions.”
• Standards/tech: INP (Input to Next Paint), LCP (Largest Contentful Paint), CLS (Cumulative Layout Shift), WAI-ARIA, WCAG (WAI = Web Accessibility Initiative), TextDecoderStream / streaming, Model Context Protocol (MCP), MCP apps (text/html mcp-app resources, URIs), OAuth.
• Concepts: taste as trainable, science-corrected pattern-matching; tacit judgment; external standards vs. internal bias; slop (2025 word of the year); perceived value → profit/purchase intent; immersion + valid cues + corrective feedback (+ humility); echo chamber; externalize a standard; cross-pollination; “perceives differently, not just judges better”; the button vs. its environment; AI harness controlling the environment; Pixar Brain Trust (“ugly babies,” diagnose-not-prescribe, no hierarchy); time to value ≠ time to completion; “bring the info to my house”; preserving the author’s brand/taste in AI UIs; Montessori-style learning by real cues.

Video: https://www.youtube.com/watch?v=VlKTUq0pdL0 — Transcript via yt-transcript.sh; outline generated from the transcript.