This week, I learned:
- I think Indigo serves Girnar Masala Chai, which I quite liked! ChatGPT
- LLMs found an archeological site! Unfortunately, it had already been discovered in 2025, so it wasn’t a totally new one, but close.
- Link lets agents control your wallet.
npx @stripe/link-cli@latest. Not available in Singapore yet #ForNow. I expect banks, retailers, and government services will soon start offering this, too. PS: I learnt this from an “Updates to Link’s Terms of Service and Privacy Policy” email. First time such an email led to such a useful discovery! - Claude’s interactive connectors and OpenAI plugin extensions operate on the same MCP Apps protocol. Good to see there’s a common standard, but I’ve yet to build an extension to see how similar they really are. ChatGPT
- … “talk through it with an LLM” is going to make complex software more accessible in the same way that GUIs did. LLM as Interface. Very true - I feel this myself. I prefer telling ChatGPT to operate my software and I believe more people will do this.
- “Read this entire project and find everything that’s broken, inconsistent, inefficient, or no longer needed. Fix what you’re sure about and list the rest for me to review.” - A useful prompt via Alex Albert. It’s worked well on a couple of uses so far (on Codex GPT 6.1 Sol High).
- Software installation instructions are increasingly being written for agents than for humans. [Siddhartha]( 27 https://www.linkedin.com/feed/update/urn:li:activity:7520673274054348800/)
Mistakes I made
- I said “US … is supposed to airdrop support within 24 hours of incidents” while discussing Singapore’s defence arrangements.
Correction: Singapore and the United States are close defence partners, but Singapore is not a US treaty ally and has no established blanket entitlement to American military intervention within 24 hours. Singapore explicitly plans its national defence without assuming that another country will rescue it. Singapore Ministry of Defence: Defence Minister’s 2025 speech · Singapore Ministry of Defence: Parliamentary reply, 13 January 2026.
MEDIUM · UNSUPPORTED - I said “Mansa Musa was … 12th century” while comparing the wealth of historical rulers with modern billionaires.
Correction: Mansa Musa ruled the Mali Empire approximately from 1312 to 1337, in the 14th century, not the 12th. His famous pilgrimage to Mecca occurred in 1324. National Geographic: Mansa Musa.
LOW · FALSE - I said “Stalin, Mansa Musa, Augustus Caesar … all had even more wealth” while comparing historical rulers with modern billionaires.
Correction: Historical rulers controlled enormous state resources, but political control is not equivalent to personally owned wealth. In particular, Soviet state assets cannot simply be counted as Stalin’s personal fortune. Mansa Musa’s wealth is also impossible to quantify reliably using modern net-worth calculations. Comparisons with contemporary billionaires must distinguish personal assets from control over an economy. University of Warwick: The USSR State Budget under Late Stalinism · History: Mansa Musa’s wealth.
MEDIUM · OVERSTATED - I said “you run it locally, which is way too expensive … Google … charge you per token. That is cheaper” while comparing local inference with managed model APIs.
Correction: Managed inference APIs are often cheaper for low or unpredictable usage, but self-hosted inference can be cheaper for sustained, high-volume workloads with sufficient GPU utilization. The correct comparison depends on model, throughput, hardware, utilization, operations costs and data constraints. Google explicitly describes self-hosting on GKE as cost-optimized for predictable, high-volume workloads. Google Cloud: Choosing agentic AI architecture components.
MEDIUM · OVERSTATED - I said “in an emergency room, you can only deploy, at best, a local model” while discussing AI-assisted annotation of medical procedures.
Correction: AI systems in emergency departments can be deployed locally, in a compliant private-cloud environment, or through appropriately secured managed cloud services. The permissible architecture depends on applicable healthcare regulations, contractual safeguards, latency, reliability, connectivity and patient-data requirements. For example, US HIPAA rules permit cloud processing of protected health information when the required agreements and safeguards are in place. US Department of Health and Human Services: HIPAA and cloud computing.
MEDIUM · FALSE - I said “LLMs … respond to surveys … the same way … 80% of the time” while explaining an experiment using synthetic respondents.
Correction: There is no established general 80% accuracy rate for predicting individuals’ survey responses from demographics. Stanford’s research found that agents grounded in extensive interviews achieved approximately 85% of participants’ own two-week test-retest consistency, which is not 85% absolute accuracy; demographics-only agents performed worse. More recent research finds that synthetic respondents can substantially misestimate public opinion, especially for demographic subgroups. These systems can support ideation and hypothesis generation but cannot be assumed to substitute reliably for human surveys. Stanford: Generative Agent Simulations · Pew Research Center, September 2026.
HIGH · UNSUPPORTED - I said “just around then, Gemini released a model that could handle 1 million tokens” while describing a document-conversion prototype developed between January and March 2026.
Correction: Google first announced Gemini 1.5 Pro’s million-token context window on 15 February 2024 and made it more widely available in May 2024, nearly two years before the stated project timeline. The availability of a larger context window also does not, by itself, make a document-conversion pipeline obsolete; conversion fidelity, formatting, cost and reliability must be compared. Google: Gemini 1.5 announcement · Google: May 2024 developer release.
MEDIUM · FALSE - I said “Opus 4.6 Thinking … became as smart as a professor” and “you have a professor in your pocket” while describing AI intelligence using IQ-test results.
Correction: AI models can match or outperform highly educated humans on particular reasoning or knowledge tests, but these scores do not establish equivalence to a professor’s overall intelligence or professional capabilities. IQ tests are normed on humans and do not automatically provide valid cross-species-or-system comparisons, particularly for models with uneven verbal, numerical and visual abilities. I should report specific model scores against stated human comparison groups on particular tests, rather than translate benchmark scores into academic qualifications. OECD: Assessing AI with tests designed for humans · Research: Evaluating LLM intelligence using IQ tests.
MEDIUM · OVERSTATED - I said “cost per million tokens … the entire Harry Potter set of books, all seven” while explaining model input-token costs to students.
Correction: The seven Harry Potter novels contain approximately 1.1 million words, which translates to roughly 1.5 million tokens using standard English tokenization estimates, not one million tokens. A one-million-token price therefore understates the cost of reading the complete series. Actual costs also depend on the model’s tokenizer, input pricing, caching and generated output. Wizarding World: Official seven-book word count · OpenAI: Understanding and counting tokens.
LOW · FALSE - I said “worlds obeying the laws of physics” while demonstrating AI-generated interactive environments and suggesting applications to manufacturing simulation.
Correction: Generative world models can produce interactive environments with visually plausible physical behaviour, but they do not guarantee physically accurate dynamics or engineering-grade simulations. For manufacturing decisions, their outputs must be validated against established physics engines, engineering models or experimental measurements. Google DeepMind itself acknowledges significant limitations in world-model accuracy, interactions and consistency. Google DeepMind: Genie 3 capabilities and limitations.
MEDIUM · OVERSTATED - I said “the Nobel Prize for Chemistry … went to the discovery of this particular molecule, ATP synthase” while demonstrating an interactive molecular visualization.
Correction: The 1997 Nobel Prize in Chemistry recognized Paul Boyer and John Walker for elucidating the enzymatic mechanism of ATP synthesis, including the functioning and structure of ATP synthase, not for discovering the molecule itself. The other half of the prize went to Jens Skou for discovering the sodium-potassium ATPase, a different enzyme. Nobel Prize: Chemistry 1997.
LOW · FALSE - I said “that is what creates the energy” while explaining how ATP synthase works in a molecular visualization.
Correction: ATP synthase does not create energy. It uses energy stored in an electrochemical proton gradient to drive the synthesis of ATP from ADP and inorganic phosphate. This converts energy from one form into another, storing it in ATP’s chemical bonds. The proton gradient is established by other processes, such as mitochondrial electron transport or photosynthesis. NCBI Bookshelf: The Mechanism of Oxidative Phosphorylation · Science: Structure and mechanism of ATP synthase.
LOW · FALSE
Questions I was asked
- Question: How do I turn an AI-generated sports highlights video into an engaging story? Asked after showing a montage with factual captions, music, and selected American football clips.
Answer: Start with the story, not the clips. “A child carrying jam encounters an animal” has more information, but “Little Red Riding Hood meets the big bad wolf” tells a story. Make the captions carry the narrative, then select scenes and music to support it. - Question: Do threats work better than rewards when prompting LLMs? Asked while I explained behavioral benchmarks behind my LLM Psychologist title.
Answer: No. The stick is much worse. In my experiments, threatening models worsened performance for roughly 80% of them; incentives also hurt most models, but somewhat less. Smarter models are less affected. Test prompting advice against actual models and problems instead of assuming human psychology applies. - Question: Are students increasingly worried about their job prospects because of AI? Asked while comparing enterprise AI adoption with what students were experiencing in the employment market.
Answer: Not structurally, from what I see. Students worried about placements before AI, during COVID, and during economic downturns. AI gives that anxiety a new anchor. The actual employment evidence is mixed, and I don’t have enough data to confidently attribute changes to AI. - Question: Should I build a Cloudflare connector for AI agents? Asked after building a Flutter CRM in six hours without writing code and considering what infrastructure to integrate next.
Answer: No. Cloudflare already has excellent connectors. I disabled its MCP because the CLI was so good. Don’t compete with people who are already doing an excellent job. Compete with the rubbish. SAP, Darwinbox, Concur—making clunky enterprise software accessible through agents is worth exploring. - Question: Can small niche businesses still survive when AI giants can do so much? Asked while considering whether independent businesses could compete with increasingly capable frontier agents.
Answer: Yes. Small teams can now do enormous things. But intelligence and domain knowledge are becoming cheap, so they’re weaker defenses. Build around assets AI cannot easily copy: customer trust, relationships, exclusive resources, licenses, regulation, and reputation. - Question: But shouldn’t students at least learn to verify what agents produce? Asked after I described designing TDS around capabilities agents cannot already perform.
Answer: If an agent can verify it, there’s no point learning that either. Test whether agents can perform the verification, not just the original task. That boundary keeps shifting, so what we teach must shift with it. - Question: Should countries accelerate AI development or slow down when governments disagree about its risks? Asked at the opening of an SUTD guest lecture about conflicting national approaches to AI governance.
Answer: I don’t spend much time deciding what the world should do. I’m a tiny cog and cannot meaningfully control the global race. I can control how quickly I learn and use AI, so I’m going fast. That doesn’t make the safety question unimportant; it’s a question I don’t have much influence over. - Question: What should programmers learn to stay competitive? Asked through the audience’s questions after seeing an agent complete a week-long data-science examination in minutes.
Answer: Programming is not what programmers need to learn. Maybe knowing what to program is important. Play with lots of programs, discover which ideas work and which don’t. Stop programming. Continue creating programs like crazy. - Question: Which industries offer the best opportunities for students learning data science? Asked through the audience’s questions about careers as AI becomes better at conventional data-science work.
Answer: Look for industries where data is heavily underutilized, especially physical data. I’d bet on manufacturing and logistics over the next few years. Banking and other established digital sectors already use a lot of their easy-to-access data. Video, audio, and telemetry are harder to collect and have more room for growth. - Question: How do we demonstrate that students have actually learned something in an AI-native course? Asked while planning to use videos and repeated assessments as evidence of learning for academic stakeholders.
Answer: Don’t assume measurable learning is the only measure of course effectiveness. If a skill is measurable and trainable, an agent may learn it too. Broaden what counts: discovery, exposure, judgment, and changes in how students approach problems. Mine their attempts and discussions for evidence, including things that resist easy quantification. - Question: Should we create a centralized skills repository so future instructors can reproduce our AI-generated assessments? Asked while planning for course continuity if new teaching assistants or instructors take over TDS.
Answer: Maybe, but the material already exists across our repositories. First point good agents at those repositories and see what they cannot figure out. That tells us what actually needs reorganizing. Don’t build another repository just because the existing material looks scattered. - Question: Should we scale a container-based interactive learning application into our main exam portal? Asked after proposing a KodeKloud-style interface with runnable files, exercises, and agent-generated questions.
Answer: Don’t worry about scaling. Put one question into GA5 and experiment. Don’t productionize it for a year. Discover what the right learning experience and assessment design should be; by then agents will probably rebuild the software cheaply. Knowing what to build matters more than preserving what you built. - Question: What does “review” mean in specifying, reviewing, and verifying agent work? Asked while I explained techniques extracted from my ChatGPT conversations for managing agent workflows.
Answer: Review is how I change the agent’s direction after seeing what it produces. I may reframe the question, replace assumptions, choose among alternatives, correct what I meant, or push an agent past its hesitation. It’s not just checking progress; it’s steering the work. - Question: What should we tell students who ask what they’re learning and what we’ll teach in live sessions? Asked during TDS course ideation after students complained that unfamiliar assessments didn’t identify specific tools or skills.
Answer: Tell them to figure it out. Learning is a choice; one question can teach different people different things. We create an environment in which they can learn, not a fixed list of facts to transmit. They can ask an agent to infer the intended learning. Part of the lesson is learning to deal with something they’ve never seen before.