Get the job done — and verify it.

The orientation began in the wrong room

Before anyone said a word about data science, there was a mix-up. The email had one meeting link; the calendar invite had another. About seven students were sitting in the wrong call — and so, for a while, was the instructor. 03:31

Nobody designed that. But it was a fair preview of a course that, as the next hour made clear, sometimes does design exactly that: incomplete instructions, a second channel you were supposed to notice, a problem that is solved by whoever figures out where the real room is.

JK was only half joking. For four or five terms, he said, the complaint had been the same: TDS is very hard, and it doesn’t feel like it is testing what students know. What followed was not five minutes on hardness. It was seventy minutes of students, one after another, pushing on the course’s philosophy to see whether it held.

She scored 34.5. She couldn’t say what she’d learned.

Twenty minutes in, Sushmitha unmuted. She had joined late and wasn’t sure if her question had already been covered. She had attempted GA0, the course’s first graded assignment — 25 questions worth 35.5 marks — and scored 34.5. Nearly perfect. And that was the problem.

Here is the paradox in its purest form. When agents do most of the visible work, learning becomes hard to see from the inside. The score says one thing; the student’s own introspection says another. Which one is lying?

Anand didn’t answer the question she asked. He answered a phrase she had used in passing.

His argument: if you couldn’t do something before, kept failing, and can do it now, something changed — even if you can’t name it. “I am less worried about you not being able to name what you have learned,” he said, “but I know that you have learned because you were not able to before and you are able to now.”

Then he reached for an analogy from forty years ago.

He offered a few of the words that are starting to appear 25:00 — and admitted they were provisional:

It is a generous answer, and a slightly risky one. A score after many failures is evidence that something in the workflow changed — not proof of what changed, or that it will transfer. Later in the call, another student, Sairam, would mention that he too had hit 34.5 — and watched the portal reset him to 28.5. Scores, it turns out, are also things you have to verify. The rest of the session was, in effect, a long negotiation over Sushmitha’s question: if the agent did it, what exactly is mine?

A course that rewrites itself every term

JK’s explanation for the difficulty was structural. “Every term has been a completely new version of TDS,” he said 05:39, tracking a field where “the practical aspects of the tools in data science have become more agentic.” The programme’s ask: embrace “this agileness that is happening outside,” because not adapting to it “will hurt you in the future.”

Anand put it more bluntly. This wasn’t true three months ago, when agents were less capable, he said; it certainly wasn’t true six months ago.

Doubly hard because the skill is tough, and because the usual escape hatch — copying from seniors — keeps closing. (“Copying is a good thing, by the way — please copy,” he added. The course just keeps making it less useful.)

Why everyone is tired by lunch

Anand opened with something JK had mentioned earlier: these days, both of them felt tired by the afternoon, not the evening 06:44. His explanation came from a TEDx talk by Kristina Kallas, Estonia’s education minister, framed through Bloom’s taxonomy: LLMs and agents are taking away the bottom layer.

He went further — that harder thinking literally burns more of the brain’s energy, which is why the tiredness arrives early. Treat that as a working metaphor rather than settled physiology: a 2025 review in Trends in Cognitive Sciences describes the mechanisms of cognitive fatigue as still debated, with both metabolic and motivational accounts in play1. JK later sharpened the point in a way that matters more for students:

Anand’s parting advice to one student was the same idea in a sentence: “Use AI more to the point where you get tired. If you are getting tired, you are progressing.” 37:05 It’s a good motivational rule. It is not, by itself, a good diagnostic — and that distinction deserves its own page.

Ignore the videos. Solve the problems.

The second principle turns the usual classroom upside down. There is course content — the teaching assistants prepare it — but it is not where learning is supposed to start.

A student in the chat tested it directly: how important are the course videos on YouTube? The answer was as blunt as anything said all evening 25:59: “No content is important unless you have a need for it… ignore all the videos. Solve the problems.” Watch one of Anand’s videos for a minute; if it isn’t helping, toss it. Try five sources; when the sixth works, “now you’ve discovered a new source of learning.”

The philosophy, written down — and operationalised

Load live page here

Load live page here

There is an obvious tension here, and it is worth naming. Someone is still writing all that content — and it is still useful when a problem exposes a gap. The claim isn’t that content is worthless; it is that content is a reference you pull, not a syllabus you push through. That matters most for the one exam that is different.

The end-term, Anand explained, is “the only assessment where you don’t have access to the internet,” your classmates, or most tools 30:22. It tests what can only be tested without an agent: “do you know how to prompt, do you know how to verify?” Asked by Biplab whether it would draw on graded-assignment questions or the recorded videos, his answer was “most likely neither” — the questions are application-based, and the preparation is the practice:

What 180 attempts at GA0 look like

If the exams are the curriculum, GA0 is the first lesson. It is open from 21 September to 11 October, and it spans browser devtools, GitHub, file wrangling, APIs, deployment, local models, evaluation rubrics, property-based testing and a network-game detective puzzle. Here is how the public attempts looked the day after the orientation.

The hardest question — solved by fewer than three in ten — is the one most like a real investigation: probe a transaction graph with a limited number of queries to identify the compromised account and the shortest proof path. No amount of pasting it into a chatbot solves it. You have to decide what to ask, in what order, and how to know when you’re done. Next hardest: getting a local model running with Ollama, and deploying an endpoint. The two easiest — at 78% — were computing a variance and hunting a bug with property-based tests. Execution is cheap. Setting up the environment, and knowing what to check, is not.

The agent executes. Everything else is yours.

“Use agents,” Anand said 13:24 — “Codex, Claude Code, GitHub Copilot, ChatGPT, Perplexity, what have you. The better the agent, the better you’ll be able to get the job done.” Earlier he had been explicit that this is a shift for the course itself: “agents are able to handle the tools, so start using agents even more than the tools, though an awareness of the tools helps.” 09:12

But handing over execution doesn’t shrink the human job. It moves it up.

The last column is the one that rewrites the course. “How do we know that the agent’s gotten it right?” Anand asked 10:30. “How do we know the people have given you the right answer? Do we know how to ask the right questions to be able to verify it?” His summary of the ethos: “get the job done somehow correctly.”

There is a reason this is now the focus. In June, Anand and a student pointed coding agents at a TDS-style exam designed to resist AI and walked away; they scored 9/10 and 10/10. If an agent can do the execution unattended, a course that only tests execution tests nothing. What’s left is ownership — which Anand called one of the few skills likely to survive “even after AI takes over your job”:

“What is 1 + 1?” is not the question

Principle four is the one that makes students angriest, and Anand explained it with a deliberately mischievous example 15:37. Try it.

“If you type ‘2’ in the answer, it will say wrong. You’ll say, ‘Wait, the question says 1+1!’ Yes, but there is a hidden piece of text.” The point isn’t the trick. It is that real requests arrive the same way:

So the course says it outright: you will be given “incomplete information… confusing information… literally wrong information,” and the problems are still “broadly solvable.” 16:21 The skill being trained is skepticism about the obvious specification — “both when it’s coming from AI as well as coming from anyone in this course.” That includes the instructor. It also explains GA0’s strangest line: “It’s hackable… Hacking is allowed.” Reading the page’s source is a way of reading between the lines.

“Should we concentrate on the how?” — “If you have time.”

Sushmitha’s worry came back in three more forms. Rakshana asked whether students should study how the LLM was approaching each problem 27:08. Biplab asked whether they should be able to explain their solutions 47:20. Divyansh asked whether, after getting the right answer, they should go back and ask the LLM how it got there 53:42.

First, Anand reframed what is being tested. “I am testing the competencies that the industry will be expecting from graduates” — will, not does. By his rough estimate, only 2–5% of companies have caught on, the ones hiring forward-deployed engineers; more will by the time this cohort graduates 27:40. (That’s an impression from his vantage point, not a labour-market statistic.) Then he explained why the course is so crowded:

He was candid about where students’ heads would be in a few weeks: “I just badly want the marks, I don’t give a damn how it gets done.” 54:10 So, he advised, don’t add a second worry to the first. “Two beatings on the head becomes a little harder — just relax in one beating.” 54:36

Then Carlton, from the course team, stepped in with the evening’s sharpest counterweight.

Carlton’s point connects straight back to the third verb. A result that “came in right for the marks” is not a result you can put “into a production system.” Understanding is what lets you trust the output. And Anand, it turns out, agrees more than his slogans suggest — the exercises are being built to force exactly this:

The synthesis is not “never learn internals.” It is narrower and more useful: let agents absorb the syntax and routine execution; learn enough structure to notice, diagnose and verify when they fail. The understanding arrives through the failures the course plants — which is why Anand could say, without contradiction, that it “will feel like the agent is doing all the work.”

The lesson nobody asked for

The most instructive moment of the evening was an accident. Vasumathi, a chemistry teacher from a non-computer-science background “learning for learning’s sake,” had two questions. Here is what happened, as the transcript recorded it.

He was answering a question she hadn’t quite asked, and said so. But it is the cleanest demonstration of context engineering anyone could have staged. Signals come from many sources; the human decides which ones reach the agent. “The broader the signal, the more data that gets into your context and therefore the agent’s context, and the more problems they are able to solve.” He promised questions built the same way: you’ll see one or two things, and miss others “because you didn’t hear, because you didn’t read.”

Then, with transcripts on, he scrolled back to her first question 34:30 — and answered it with a story about a colleague, Abhishek, who had built a screenshot-to-clipboard app.

“Is that fair?”

Carlton read out the hardest question of the night from the Q&A panel 39:21: “Will students who do not have access to costly models be disadvantaged compared to those that do? Is that fair?”

He meant it as realism, not indifference. A student from an underprivileged background has less to eat, fewer industry connections, less parental support, less time for homework. At work, one person arrives with a splitting headache after an argument at home and another has had a fantastic day; “I have a 2-lakh loan; the other person has an ancestral home” 40:23 — and both get the same assessment. Model access is one more uneven starting point. The skill is deciding “where do I spend the scarce resources that I have?” 41:00 — less on a model and more on friends, or on your own engineering, or on squeezing everything out of “$20 or 499 rupees.”

JK added the resource that costs nothing 42:36:

The question returned, smaller, at the very end. Divyansh asked which AI was best to pay for 74:52. JK answered with a counter-question: “If you were supposed to buy a bike that is most suited to you, which bike will you select?” You research it. Each of the three, he noted, had their own pet preference — Anand’s being, in JK’s words, “maybe Luna plus Sol.” 75:32

Humans are tools too

The phrase sounds harsh, and Anand pre-empted that 17:18: “I don’t mean that in any derogatory way; I mean that they have capabilities that are very complementary to agents.” With agents taking over a big part of the job, “leveraging the knowledge of other people is something that humans are still slightly better than agents at.” 09:52

So the course doesn’t just tolerate collaboration; it engineers for it. “If you feel copying is bad, think of it as group work,” he said 11:39 — and warned that “finding a friend is much tougher than actually doing the work yourself.” Some tasks are explicitly designed so that you have to.

When Carlton relayed a question about which skills complement AI 43:16, Anand shared his screen and an evolving list — “literally changing every week” — sorted into four tiers.

Note where verification lands on the list: not in the tier that’s vanishing, but in the one that’s growing. Accountability is the reason the third verb exists. If your neck is on the line, “it worked when I tried it” is not good enough.

“Should we care, or just blindly trust?”

Shivam had been pasting questions — including roll numbers and personal details — into LLMs, sometimes as screenshots 56:27. “This is a security-related issue, so we have to care about it or just blindly trust on that?”

“Your choice,” Anand said, and confessed his own: his bank passwords already sit with Dropbox and Google, while a few things live only on his own machine and a USB stick — with the risk of losing them. OpenAI and Anthropic are just more entities on that spectrum 57:08.

Biplab pushed for specifics: “How long do they store the data and what is the possibility that they may misuse this data?” 63:58 Anand gave his rough personal rule — and, tellingly, said the accurate answer is something “you should obviously ask a good agent to solve.” So we did the checking.

“The portal is lousy.” So change it.

Sairam, a product manager who evaluates products for a living, delivered the evening’s most candid review 59:59. He loved the questions. He had spent “18-plus hours” on the portal. And then:

Navigating 25 questions back and forth, he said, was “taking the precious cycle away from solving the problem.” His request: an agent that could reorganise the site. It is exactly the kind of friction the hardness counterpoint warned about — and the team’s responses turned it into a lesson anyway.

Anand asked for something specific: “please see if you can reproduce bugs with logs.” Replays — which agents can now help produce — let the team match a report against the Cloudflare logs and the codebase. If you can do that, “that would be the most valuable contribution you can make to this course.” 61:46 Report them on Discourse or GitHub issues, he told Dharitha later, with full diagnostics 68:06.

Carlton pushed back on the rest 62:43: there is already a way to jump to any question, and the score resets are “kind of by design… you may agree or disagree, but that’s how it is.” Some questions are built so that answering one breaks another unless you have two servers running at once. And the all-on-one-page layout stays, Anand said, because “the ability to pick the question to solve at a priority is also an important skill to learn.” 69:19

// A bookmarklet is just a URL that runs JavaScript
javascript:(() => {
  const links = [...document.links]
    .map(a => a.href)
    .join("\n");
  navigator.clipboard.writeText(links);
  alert(document.links.length + " links copied");
})();

Then came the line that captures the whole principle — floated, and immediately withdrawn: “It may make sense to worsen the exam portal so that you can improve it. We won’t do that, but it’s a thought.” 71:27

It matters because initiative is what Anand says he actually looks for. Early on, introducing principle five, he was unusually personal about it:

When Yasin asked how he would spot the skill of completing the objective, the answer had two parts 66:57: “solve the toughest problems. As simple as that” — and, complementing it, “what are people doing that nobody asked them to do?” A student who reproduces a portal bug with logs, or ships a bookmarklet that fixes the navigation for everyone, has answered both.

The instructors aren’t keeping up either

The most disarming admission came early 17:51. That morning, Carlton had told Anand he felt he wasn’t keeping up with the course. “Yes, I’m having exactly the same phenomenon,” Anand replied. “So, you are not alone.” New students, he argued, may even have it easier — they have nothing to unlearn. For now.

Here is the deepest tension in the course. It teaches current tools, because you can’t practise on abstractions — GA0 is full of FastAPI, GitHub Actions and Ollama. But the tools have a half-life of months, so the most durable thing it can leave behind is the habit of relearning. The course team is trying to model that in public: this term, students will watch the team design new questions in live sessions — “we’ll do, you watch,” Anand said, with “some say” for students 61:22.

“What the college tries to teach us?”

Near the end, Siddharth asked the question everyone had been circling: “Sir, by this course, what the college tries to teach us? I don’t understand basically.” 51:11

Anand declined to give him an answer, and gave him a method instead.

“It’s okay that you don’t understand the answer, because I haven’t given you the answer; I’ve given you the approach.” (This page is a small instance of the method: a recording, a transcript, agents drafting — and then a human-directed pass checking every quote against the transcript and every claim against its sources.)

Then Lavanya, who had been listening since 5:30, offered the summary the room had been building toward 52:06:

Lavanya admitted she’d been worried — doing well, solving the questions, but wondering, like the colleague who didn’t know his own app’s language, whether that counted. Anand’s reassurance was characteristically double-edged: “Don’t worry, we’ll make the course much tougher until you actually do get it.” 53:32 “Oh, thanks for that also, sir,” she said.

Notice the clause Anand added to her summary. Lavanya’s version ends at the goal is reached. His ends at make sure you’ve verified it. That last clause is the whole difference between delegating and abdicating — and it is the answer to Sushmitha, too. What is hers, when the agent did the work, is the specification, the context, the noticing, the checking, and the willingness to put her name on it.