machine-learning on S Anand

Things I Learned - 14 Dec 2025

Sun, 14 Dec 2025 00:00:00 +0000

This week, I learned:

Zillow Offers, the company’s “iBuying” arm, which was shut down in November 2021 after losing hundreds of millions of dollars. The core failure was not just an algorithmic error, but a fundamental misunderstanding of the limits of machine learning in high-stakes, low-frequency trading environments like real estate. Zillow relied on its “Zestimate” algorithm to predict future home prices and make instant cash offers, but the model failed to accurately account for real-time market volatility and “adverse selection”—savvy homeowners sold their properties to Zillow when the algorithm overvalued them, but kept them when the algorithm undervalued them. This left Zillow holding thousands of homes it had overpaid for and could not profitably resell, forcing a $304 million write-down and the layoff of 25% of its workforce. Zillow Q3 2021 Shareholder Letter (PDF) #
There’re a good number of AI insurance products in the market. #
- Munich Re aiSure - for AI vendors and companies deploying AI; can cover business losses (like lost revenue / business interruption) and legal damages when AI performance errors (incl. hallucinations) cause harm.
- Munich Re aiSelf - for teams using self-built or bought ML models; helps cover the financial downside when models underperform or drift over time.
- Munich Re aiSure - General Liability - covers damages and financial losses from lawsuits (e.g., claims that AI decisions were biased/discriminatory).
- Armilla Insured (AI Liability Insurance) - affirmative AI liability cover (Lloyd’s coverholder; partners include Chaucer) that can cover legal defense costs, settlements, and third-party claims when an AI model underperforms.
- Armilla + Chaucer standalone AI liability (announcement) - focused on “mechanical underperformance” (incl. hallucinations and model drift) and the liability that follows.
- AXA XL GenAI Endorsement for CyberRiskConnect - add-on to cyber insurance for companies building their own GenAI; covers things like data poisoning, copyright/usage-rights mistakes, and AI-regulatory violations.
- Coalition Affirmative AI Endorsement - clarifies cyber coverage applies when AI causes a security failure, and extends funds-transfer-fraud triggers to deepfake-based instructions.
- Coalition Deepfake Response Endorsement - adds response support for deepfake incidents (technical analysis + legal + reputational help), not just “classic hacking.”
- Tokio Marine Kiln Technology Errors & Omissions - tech E&O with generative AI coverage available by endorsement (aimed at software/SaaS/tech services).
- Tokio Marine Kiln Cyber Ctrl suite - cyber/tech cover where AI-related add-ons can include AI regulatory proceedings, data contamination, and “LLM hijacking.”
- Hiscox Technology PI (UK) - AI clause - explicitly covers client claims arising from your use of AI (incl. genAI) as part of the services you deliver.
A key lesson from Who Validates the Validators is that we learn our preferences as we evaluate. So make it cheap to evaluate (create outputs) AND cheap to revise criteria.
Cookies taste wonderful when eaten hot.
⭐ Constraints as opportunities. On long flights, I read more since I’m less distracted by guilt (“Should I answer email or code instead of wasting time?”) or FOMO (“Let’s click that link”) since I have no choice. Setting aside “quiet time” doesn’t work as well, since I have more choice. This constraint (no Internet) became an opportunity (reading time). I knew this before-hand, but had to experience it to appreciate it, and acknowledge it consciously to realize it. That takes repeated (2+) trials and reflection. A workflow to convert constraints to opportunities could be:
- List constraints. (Like fish in water, we aren’t used to thinking of constraints as constraints. Also, this means more constraints => more latent opportunities!)
- List opportunities they offer. (Creative prompting helps; reflecting on the answers helps more.)
- Try any 2+ times. (Gives room to settle in.)
- Document learnings. (Explicit reflection is better than implicit awareness.)
Notes from Thoughtworks Radar, Apr 2025
- Architecture advice beats architecture review. Architecture Review Boards hinder workflow. An architectural advice process (anyone makes architectural decisions, taking advice from experts, logging in Architecture Decision Records) works better.
VectorChord is a faster pgvector alternative.
“Learning is not the product of teaching. Learning is the product of the activity of learners.” – John Holt
Music labels never became streaming platforms themselves. The real money is in concerts. Streaming just makes you famous enough to book gigs. But movies/TV shows are far more expensive to produce than music. So streaming platforms invest in content (Netflix, Apple) and studios stream (HBO, Disney) Claude
Notes from Better Ways to Build Self-Improving AI Agents
Quotes from Life is more than an engineering problem, interview with author Ted Chiang.
- Magic is intent-centric. “Magic means that … the universe responds to your intentions in a way that the laws of physics as we understand them don’t.”
- LLM reasoning is a weak analogy. “My liver was running this old program, but all I needed to do was update the software and now my liver is functioning much better, even though the hardware is the same.” No one says that. It’s not a useful way of thinking about the liver, and it is not a useful way of thinking about the brain either.
- Art won’t die. Art is all about context. It’s not an activity like tightening bolts, where I don’t really care whether someone used a conventional wrench or a pneumatic wrench, as long as the bolts are tight.
- Alignment may not happen. When corporations behave badly, should we consider that an alignment problem? But why do large corporations behave so much worse than most of the people who work for them? And could that be fixed by solving a math problem? I don’t think so.
- LLM relationships are different from human. … people have their own preferences, while things do not; you do them harm because you are ignoring their preferences. (Companies) might create the illusion that AI systems have preferences.
- .. it’s theoretically possible for us to build digital entities that have subjective experience.
Notes from Developing our position on AI by Recurse Center:
- Learning happens at the edge of competence. AI has a moving jagged edge, so constantly re-try your impossibility list.
- Learning happens on what you care about. Use AI to expand your agency (by complementing or deepening), not replace it.
- Learning generously means being open to different perspectives, without judgement or dogma. Try new perspectives.
⭐ ‘We tested one of the most common prompting techniques: giving the AI a persona to make it more accurate We found that telling the AI “you are a great physicist” doesn’t make it significantly more accurate at answering physics questions, nor does “you are a lawyer” make it worse. This doesn’t mean that personas can’t be useful - for example, they change how the AI answers questions, the format of output, and maybe other factors as well.’ Prompting Science REport 4: Playing Pretend: Expert Personas Don’t Improve Factual Accuracy
If YouTube embeds fail with an “Error 153 View player configuration error”, it’s because the server probably has a Referer-Policy: same-origin and needs to switch to Referer-Policy: strict-origin-when-cross-origin. Simon Willison
Adding a [dependency-groups] section to pyproject.toml with dev = ["pytest"] ensures that pytest is automatically installed by uv because dev is a default group. Simon Willison
CloudFlare Python Workers has full Pyodide support. That means most Python apps will now run on CloudFlare Workers, with low latency worldwide. This is a big deal.
Smart contracts are programs that run on blockchains like Ethereum, e.g. to convert currencies, lend/borrow, buy NFTs, etc. These may contain bugs. Anthropic built a benchmark of real smart contracts with known bugs, had agents exploit them, and simulated $550 mn in theft. They also nade $3.5K exploiting real bugs - at a cost of $3.5K. So AI agents are currently at break-even for crypto-theft. Anthropic #
Notes from Cory Doctorow’s summary of The Reverse Centaur’s Guide to Criticizing AI: #
- When tech monopolies saturate their markets, their P/E collapses, reducing share value. This incentivizes bubbles.
- Automation blindness negates human-in-the-loop. When AI makes rare mistakes, humans don’t catch them. TSA misses guns, not water bottles.
- AI doesn’t need to do your job. The AI salesman just needs to convince your boss it can, especially senior jobs.
- Reference letters from professors used to signal value since they were hard to write, so professors would do it only for good students.
- Copyright expansion and regulation will likely benefit corporates, not labor.
- US Copyright Office making AI content non-copyrightable means corporates NEED labor. Else every AI work goes to public domain.
There is no strong evidence yet that Neuro-Linguistic Programming (NLP) works broadly (ChatGPT). Some NLP techniques help sometimes, but no more than other established techniques (goal-setting, visualization, etc.) ChatGPT
⭐ Just repeating a statement makes it seem truer because the brain finds it familiar, hence easier to process. This seems well-established research. The Truth about Truth
PGlite is a WASM-based Postgres implementation. It’s ~3MB. You can embed it in the browser, NodeJS, Deno, etc. It has plugin support, including pgvector.
Pejoration is when words acquire negative connotations. Euphimism escalation is another term for it. Third World → developing countries → emerging markets → Global South. Old → elderly → senior citizen → older adult. Lunatic → insane → mentally ill → mentally challenged. Janitor → custodian → sanitation engineer → facilities maintenance specialist. The opposite is amelioration. Minister moved from servant → servant of church → government official. marshal: horse-servant → horse-officer → senior military officer. Knight: servant → armed retainer → mounted warrior → knighthood honor. # #
OmniDocBench 1.5 is a benchmark for parsing realistic PDFs. Gemini 3 Pro does well on the list among the commercial LLMs. PaddleOCR-VL (0.9B) tops the benchmarks, overall.

Things I Learned - 02 Feb 2025

Sun, 02 Feb 2025 00:00:00 +0000

This week, I learned:

You can add any content at the end of a PDF file. It’s ignored. It’s an interesting way to send additional information (or just blow up the file size if you don’t like them.)
JavaScript introduces a Temporal object that will replace the Date object.
You can use embeddings as the input to a classical ML classifier. This can improve classification a lot. Nomic
As AI software becomes more common, demand for AI product managers will grow. Also as a proportion of people in an organization. https://www.deeplearning.ai/the-batch/issue-284/
Control of chips and GPU compute is what will likely be the gameplay to control AI dominance globally. Dario Amodei
Bring LLMs to the table. One mode of collaboration is using LLMs as ACTIVE participants, i.e. they CONTRIBUTE. For example, in a video call. A workshop. A classroom. A presentation. Have the LLM provide input DIRECTLY to a group of people.
Environment shapes ambient thoughts. Working in a hospital will give you ideas about how to use LLMs in hospitals, for example. People you are working / ENGAGING with are perhaps the biggest drivers.
The cost of a cream biscuit packet in India has fallen about 25 times, i.e. about as fast as inflation, between 1981 - 2024. Effectively, the absolute price has not changed. How do I know this?
- In 1981, a cream biscuit packet cost Rs 25
- In 2025, it’s available for Rs 21
- India Inflation Calculator - a rare inflation calculator with annual inflation rates baked in - shows that Rs 25 in 1981 is equivalent to Rs 540 in 2024. That’s about 25 times more than the Rs 21 it costs today.
A WebAssembly compiler that fits in a tweet deconstructs a piece of JS that creates a tiny WebAssembly calculator. It’s a great walk-through of JavaScript compression tricks and how WebAssembly works. Simon Willison
Brandon Sanderson has a series of YouTube videos where he teaches a course on magic systems.
When using AI coding agents, CLI beats APIs. Simpler models are able to use the CLI more reliably than APIs. Simon Willison
I was exploring new business models enabled by LLMs. Here are some thoughts:
- 1. Autonomous Multi-Sided Marketplaces. AI-powered platforms coordinate complex services with minimal human oversight—think “Uber for Everything, but the platform sets pricing dynamically, schedules both supply and demand, and resolves disputes algorithmically.
- 2. Collective Intelligence Ecosystems. Communities pool data, expertise, and AI models to tackle shared problems—like an open-source “GitHub for AI, but with embedded micropayments or tokenized incentives to reward contributors whenever the models are used commercially.
- 3. Zero-Employee Companies. Fully automated software entities—legal frameworks might allow an AI to manage services, pay taxes, and sign contracts. These “companies only hire humans as needed, on-demand, for edge cases AI can’t handle.
- 4. Context-Aware Knowledge Platforms. Imagine a Wikipedia that not only retrieves static info but also tailors each page in real time to the reader’s personal context, language level, and preferences—generating content on the fly. User feedback loops train the system to improve.
- 5. Data Cooperatives / Data DAOs. Groups collectively own their data and license it to AI companies on a revenue-share basis. Individuals have a direct financial stake in how their shared data is leveraged, voting on permissible use cases.
- 6. Personalized Service Layers. Similar to GitHub’s “forking model, but for entire user experiences. Each user can clone and customize an AI service (whether it’s a personal grocery shopper or a content curator) and can share or monetize improvements with the broader network.

Managing the data deluge

Sat, 10 Feb 2007 12:00:00 +0000

Peter Norvig’s brilliant talk on Managing the Data Deluge. Among other things, he talks about how having lots of data is sometimes better than having a carefully designed algorithm.

Netflix prize

Mon, 02 Oct 2006 12:00:00 +0000

Netflix has released a sample of its customers’ movie ratings at Netflix Prize. You can download these (700 MB), create an algorithm that rates the training data, run it against the test data, and see if you can get better ratings than their algorithm. If you do, you win $1 million. (Chris Anderson explains why.)

Learning robots

Sun, 27 Feb 2005 12:00:00 +0000

Of robots that learn to walk and robots that learn to talk (well, translate, at least).