Daniel An

Director, Global AI/Copilot GTM Lead @GitHub | prev MSFT, Mozilla, Google, Barclays. Founder/CEO |

[Daniel An @Daniel_An23]

https://t.co/QYNr2QkVAe

[Daniel An @Daniel_An23]

follow up q: would you send an 18yr old to harvard for a 4yr cs degree given what you just told me? is the ROI defensible?

Yes — but almost none of the reason is the curriculum, and that's the honest answer most people dance around.

What you're actually buying at Harvard isn't the CS education. It's four things: the admission signal (the hardest filter the kid will ever pass, stapled to their name for life), the peer network (the densest concentration of future founders, researchers, and hiring managers they'll ever live down the hall from), faculty/research access, and optionality — the ability to pivot into quant, policy, a startup, law, or a PhD without explaining anything. The 11–14 courses are arguably the fourth or fifth most valuable thing on campus. If a family thinks they're paying for the courses, the deal looks bad. They're not.

The market shift makes the brand worth more, not less. This is the uncomfortable part of the data from earlier: when new-grad hiring contracts 65% and a single opening gets a thousand AI-polished applications, recruiters don't get more egalitarian — they reach for cheap filters, and pedigree is the cheapest one there is. SignalFire says "shipped work over credentials," and that's true at the margin, but the Harvard name is what gets the shipped work looked at. In a tight market, credential + portfolio beats portfolio alone. The kid the contraction actually crushes is the median CS grad from a mid-tier program with no artifacts — the exact profile driving CS's near-top new-grad unemployment numbers.

The curriculum gaps are self-repairable — for the right kid. Everything I criticized (no required teams, no shipping, systems as electives) is available to a motivated student: CS 1060, CS 2620, MIT cross-registration is officially sanctioned, T4SG and clubs, three summers of internships. The gaps hurt the passive student who does the minimum and expects the degree to convert. So the real question isn't "is Harvard worth it" — it's "is this specific 18-year-old the kind who treats requirements as a floor?" If yes, send them. If they need the institution to force-feed the right skills, my redesigned curriculum matters and Harvard's doesn't provide it — a Waterloo co-op or Georgia Tech might genuinely serve them better.

The money changes the answer less than people assume. Sticker is roughly $83–87k/year, ~$340k all-in. But Harvard's aid is unusually blunt: as of the 2025 expansion, families under ~$100k pay essentially nothing, and under ~$200k pay no tuition. So for most admitted families the real cost is far below sticker, and the ROI question mostly bites the family paying full freight in cash. Even there I'd probably still say yes — $340k against a lifetime of network and optionality is defensible — but it's now a judgment call, not a slam dunk, and "take the $340k, go to a strong state flagship, and bank the difference" is a legitimate position I wouldn't argue hard against.

Where I'd genuinely hesitate: if the kid's goal is narrowly "become a salaried software engineer," full stop. That's the one career target where the 2026 data says the entry rung is sawed off — 7.5% of postings entry-level, the junior on-ramp tasks automated. Betting $340k and four years on the single job category whose floor is dissolving is the weakest version of this trade. But that's an argument against the goal, not against Harvard — and Harvard is precisely the degree that isn't a vocational bet. It hedges across every outcome, including the ones where "software engineer" stops being a stable job title by 2032.

So: send them, on three conditions. Ship something real every year, do the internships (the resume line now does more work than the GPA), and treat the concentration requirements as the minimum bid, not the strategy. The degree still clears its cost. It just no longer clears it on autopilot — and ten years ago, it did.

[Daniel An @Daniel_An23]

I asked Fable to critique Harvard's CS curriculum versus industry needs plus come up with a redesigned 4-year course sequence ⤵️

Harvard CS vs. What the Industry Actually Wants

TLDR: Harvard's CS concentration is a rigorous theory degree with a world-class front door (CS50) and an almost empty middle. You can graduate with honors having never taken operating systems, never touched a database, never worked on a code you didn't write, never worked on a team, and never deployed anything. That was a tolerable gap in 2015, when companies expected to train juniors for 18 months. It's a serious gap in 2026, when new-grad hiring at Tech Majors is down ~65% from 2019 and the surviving entry-level jobs demand people who can supervise AI systems, verify output they didn't write, and ship — on day one.

What Harvard actually requires (2025–26)

The AB basic track is 11–14 courses: calculus, linear algebra, probability (Stat 110), then a 9-course CS core — two programming courses (CS50 → CS51/61), three "formal reasoning" courses (discrete math, computability, algorithms), one systems-tagged course, one "computation and the world" course, and four advanced electives. Honors adds CS 1240 (harder algorithms), one AI-tagged course, and a fifth elective. Thesis optional. No capstone.

The load-bearing detail: the Systems tag is satisfiable by CS 61 itself — an intro course. So the only universally guaranteed systems exposure is one sophomore-level C/C++ course. OS (CS 1610), networking (CS 1450), databases (CS 1650), distributed systems (CS 2620), and security (CS 2630) are all electives, and several are grad-numbered and irregularly offered. Harvard's own tag table lists MIT courses (6.5840 Distributed Systems, 6.1040 Software Studio) as approved substitutes — a quiet admission of the catalog gap.

Strengths — real ones, worth keeping

1. CS50 is the best onboarding funnel in higher education. ~2/3 of students arrive with zero prior CS; it also pioneered institutional AI pedagogy (the CS50 Duck, ~9.4M tutoring prompts, guardrailed against just giving answers). Most schools are still pretending AI doesn't exist; Harvard built policy and tooling in 2023.
2. The theory core is genuinely durable. Three formal-reasoning courses plus Stat 110 and real linear algebra. In an era where AI writes the code, knowing what correct and efficient look like is the part that appreciates. Complexity, probability, and proof are exactly the skills you need to verify a system you didn't build.
3. Flexibility. The tag system, double-counting, joint concentrations, and the AB/SM let strong students go deep fast. Liberal-arts breadth is underrated for exactly the "communication, problem-framing, product sense" signals NACE says employers rank highest.
4. Green shoots. CS 1200 (unified algorithms-and-limitations) is smart compression. CS 1060 "Software Engineering with Generative AI" (Spring 2025, with CI/CD and automated testing) is the right course — it's just an elective capped at 66 seats in a concentration of ~500.

Gaps, ranked by damage

1. No required team software engineering. Nothing like MIT 6.031 or a project studio. The Programming 1/2 requirement mentions "good software engineering practices" but is met by individual psets. Meanwhile the strongest current employer signal — from Canva's, Shopify's, and Meta's redesigned 2025 intervrequirement clarification, planning before prompting, scoped AI delegation, and critical review of AI output in collaborative settings. Psets are solo, spec'd, greenfield, and graded once. Real work is team-based, ambiguous, legacy, and maintained.
2. No required contact with running systems. Databases, networking, deployment, observability — all optional. The BLS category that's declining is "computer programmer" (routine coding); the growing ones are developwho own systems end-to-end and security analysts (+28.5%, fastest in the field). Harvard's reqs skill profile with the latter's theory vocabulary.
3. AI is a policy problem, not a curriculum subject. Only honors students must take one AI-tagan mean learning theory. Nothing required teaches the actual emerging job: evals, agentorchestration, verification of generated code, judgment about when AI output is "almost right but not quite" (the top frustration for 66% of developers in Stack Overflow's 2025 survey). Anthropic and Google now say most of their new code is AI-written; whatever discount you apply to those claims, the directi
4. No capstone, no shipping requirement. SignalFire's 2026 report is blunt: shipped work over graduate with a transcript and zero artifacts.
5. The elective catalog can't backfill the requirements. Even a motivated student hits irregulor distributed systems, security, and big data. A requirement that can't be scheduled isn't arequirement.
Where the industry is headed (and the honest uncertainty)
The next decade, as best the data supports: entry-level task work (boilerplate, routine debugging, unit tests) is largely automated — that's precisely what juniors were hired to do, which is why the Stanford "canaristudy found a ~13% relative employment decline for 22–25-year-olds in AI-exposed jobs while olBLS still projects +15% for software developers through 2034, security +28.5%, andJevons-paradox dynamics (cheaper software → more software) have grown headcount through every prior tooling revolution. Whether AI is a complement or a substitute this time is genuinely unresolved — Dario Amodei hasargued both sides within twelve months. My read: total demand probably holds or grows, but thesappears is "junior who translates tickets into code." The job that grows is "engineer who owns a system, directs AI labor, and is accountable for correctness." A curriculum should target the second job directly.
--- The redesigned curriculum
Design principles: keep Harvard's theory rigor (compressed, not gutted); make systems, teams, t aspirational; treat AI as a raw material to be engineered with and verified, from week one;require shipping. Every course keeps a "manual mode" component — students must demonstrate they can do the thing unassisted before they're graded on doing it 10x faster with AI. That mirrors exactly what Canva foundseparates good candidates: they can debug when the AI's solution breaks.
Year 1 — Foundations
CS 101: Computation, Code, and AI Collaboration. CS50's successor. Python and C, but git, test one — the trio practitioners have complained grads lack for a decade. Every pset done twice:once unassisted (proctored, small), once AI-assisted (large, ambitious), with a written critique of what the AI got wrong. Keeps CS50's production values and duck.
CS 102: Software Construction. Individual → pairs. A typed language (TypeScript or Java). Data structures in practice, interfaces, specs, unit/property testing, code review as a graded skill. The final third: studeninherit each other's midterm projects and extend them. First contact with code they didn't wri
CS 111: Discrete Mathematics and Proof. Essentially CS 20, unchanged. Proof is the verificatio
MATH 115 / STAT 111: Linear algebra and probability (Stat 110 is already right — steal it).
Year 2 — The machine and the theory
CS 201: Computer Systems. CS 61 / CMU 15-213 territory: C, memory, machine organization, concu earns its place more every year because this is where AI assistance is weakest and mentalmodels matter most.
CS 202: Algorithms and Their Limitations. Harvard's CS 1200 is the right compression — algorithms, complexity, and computability in one required course, with a second theory course available for the honors-inclined.Cut from three required theory courses to two (counting CS 111), reinvest the slot in systems.
CS 210: Data Systems. Required, sophomore-level, offered every semester. SQL, data modeling, te of pipelines and analytical stores. There is no modern job description without data. Harvardoffering this only as elective CS 1650 is the single cheapest gap to fix.
CS 220: Machine Learning Foundations. Required for everyone, not just honors. Supervised learning, gradient descent, transformers at a working-understanding level, and — the part almost no school requires — evaluatitrain/test discipline, benchmarks, error analysis, why models fail. Evals are the new unit tes
Year 3 — Scale, adversaries, and teams
CS 301: Operating Systems and Networks. One required course merging the OS and networking esse, virtual memory, TCP/IP, sockets, what happens when you type a URL. Deep-dive versions remainas electives.
CS 302: Distributed Systems and the Cloud. Required. Replication, consistency, consensus at a practical level, queues, caching, deployment, observability, incident response. Labs on real cloud infrastructure with recost budgets. Currently this is a grad course at Harvard; it's the daily reality of every indu
CS 310: Software Engineering Studio. The flagship, and the biggest single change. Teams of 4–6berately imperfect open-source codebase. One semester of: reading code archaeologically, fixing bugs without breaking things, code review, CI/CD, planning work, using AI agents as team members — with each student accountable for every line merged under their name, however it was produced. Graded on the diff, treviews, and a postmortem. This is Harvard's CS 1060 made mandatory and given teeth.
CS 320: Security and Adversarial Thinking. Required. Threat modeling, the OWASP classics, memo section that didn't exist five years ago: prompt injection, model exfiltration, and thesecurity properties of AI-generated code. Fastest-growing occupation in the field; currently a grad elective at Harvard.
Year 4 — Judgment and shipping
CS 410: Building with Language Models. Engineering LLM systems: context engineering, RAG, agent orchestration, fine-tuning vs. prompting tradeoffs, eval harnesses, cost/latency budgets. Half the syllabus is refresheannually — instability is a feature; learning to relearn is the meta-skill. (~90% of AI job ponot research.)
CS 490/491: Capstone (two semesters). Required. Build something with real users, or make sustaficant open-source project, or industry-partnered work. Semester two is the differentiator:maintain it. Handle real bug reports, ship v2, write the postmortem. Includes the technical-communication requirement — design docs, a public writeup, a talk. Graduates leave with a portfolio, which is now worth morthan the GPA (SignalFire: proof of shipped work over credentials).
Plus 3–4 free electives — theory depth, graphics, PL, robotics, biology + CS, whatever. The lis a strength.
The count
Fourteen required CS/math courses plus electives — heavier than Harvard's basic track (11–14 tering-school BS, achievable in 8 semesters at ~2/semester. What got cut to pay for it: onerequired theory course, and the fiction that four unconstrained "advanced electives" constitute a systems education.
What I'd bet on and what I wouldn't
High confidence: teams, legacy code, data systems, security, verification-of-AI-output, and shipping will matter more in 2035, because every hiring signal from 2024–26 already points there and none of it is new — AIjust raised the price of lacking it. Lower confidence: the specific content of CS 410. RAG andke CGI-bin scripts by 2030. That's fine; the course is really teaching "how to engineer atop afast-moving substrate," and that skill transfers. The thing I'd refuse to cut under any pressure is the theory-plus-systems core — if the industry bifurcates into "AI operators" and "engineers who understand the machine," the second group sets the wages.
---
Sources behind the claims: Harvard's requirements from https://t.co/lC0rYVCgPP and the 2025–26 College Handbook; hiring data from SignalFire's 2025/2026 State of Talent reports, the Stanford Digital Economy Lab "Canaries in the Coal Mine" study (ADP payroll data), BLS 2024–34 projections, HackerRank 2025loper surveys, NACE Job Outlook 2026, and Canva/Meta/Shopify's published interview redesigns.Happy to turn this into a shareable artifact page, or go deeper on any single course syllabus.

[Daniel An @Daniel_An23]

https://t.co/cbuVlDa6XR

[Daniel An @Daniel_An23]

@fchollet basically, be a generalist not a specialist

[Daniel An @Daniel_An23]

@chamath @grok if every enterprise customer caps spend at $200/week per SWE what is the adjusted TAM for agentic coding ?

[Daniel An @Daniel_An23]

@chamath @grok if every enterprise customer caps spend at $200/week per SWE what is the adjusted TAM for agentic coding ?

[Daniel An @Daniel_An23]

for the first time in my life I’m so bullish on IBM stock

[Jessica Lessin @Jessicalessin]

Here are two things I don't understand about the AI market right now.

1) Open Source is getting very good and the hot topic but seems like Anthropic (and probably OpenAI, but I have less visibility) are ripping in revenue.

2) All anyone can talk about is the unlimited need for more data centers but then the providers are striking rent-back deals and pursuing increasingly complicated hedges.

[Daniel An @Daniel_An23]

@StringerViral Yes this is my video feel free to share

[Daniel An @Daniel_An23]

England fans belting out God Save the King , electric atmosphere in Atlanta #fifaworldcup26 🏴󠁧󠁢󠁥󠁮󠁧󠁿 vs 🇨🇩

[Daniel An @Daniel_An23]

@StoryfulNews Yes go ahead 👍

[Daniel An @Daniel_An23]

This is what football means to the world. Can u imagine the WH issuing a statement like this if the USMNT loses against Bosnia??
This is the difference between soccer vs football
President of South Korea:

[World Cup HQ @WorldCup26HQ]

KBS (news channel of South Korea) blurred the face of head coach Hong Myung-bo after their defeat to South Africa. https://t.co/1E3VVwJcf2

[Daniel An @Daniel_An23]

the thing about "migrating" from one LLM to another: it's easy. a few clicks + a few lines of code

whereas swapping hyperscaler A for B is a full rebuild

we knew the lock-in was never the model. it's your prompts/evals/ tool-calling which don't transfer cleanly

[Lindy @getlindy]

https://t.co/LdMW8nkU6o

[Daniel An @Daniel_An23]

electric atmosphere in Dallas at the
Fifa World Cup 2026 🇯🇵 vs 🇸🇪

[Daniel An @Daniel_An23]

Anthropic is the new “AI Operating System” for knowledge work

I wonder if they have hardware ambitions or if this will be left to Cupertino, OpenAi et al …

[Daniel An @Daniel_An23]

I have worked remotely since 2021 and my mental health has literally never been better

[Natalia Emanuel @NataliaHEmanuel]

Published an op ed in @nytimes: remote work hurts mental health.

Let me explain why
🧵⤵️

[Daniel An @Daniel_An23]

So far Siri AI is pretty good. 1.2b Apple users (sans EU/china) are about to realize they will never need to install/update other chatbot apps

[Daniel An @Daniel_An23]

@BibimBallers as a lifelong barca fan this would be amazing. Pedri, Frankie and Lee Lang-in in midfield🤩

[Daniel An @Daniel_An23]

As a Norwegian, I need to get this off my chest 🇳🇴

English alphabet: 26 letters
Norwegian alphabet: 29 letters (includes Æ, Ø, and Å)

English spelling: Haaland <> Odegaard
Norwegian spelling: Håland <> Ødegård

"å" = "aw/oh" not "aa"
"ø" = "uh" not "oh"

1.2B world cup viewers mispronouncing Håland & Ødegård🤣 #FIFAWORLD26

[Daniel An @Daniel_An23]

https://t.co/aHkGCkXpEu

[Andrew Curran @AndrewCurran_]

Microsoft has fine-tuned a version of DeepSeek and is considering it as a lower-cost alternative to OpenAI and Anthropic models for Copilot Cowork. Huge development considering the history of the OpenAI/MS relationship.

[Daniel An @Daniel_An23]

every 'expert' pre IPO: this is institutional selling to retail. stock will tank. do not touch this stock!!
what is actually happening: billionaires buying $1B in stock on IPO day. stock up 26% since Friday