Jun 26 | What Is Intelligence? by Blaise Agüera y Arcas

Currently reading. First impression: thoughtful, dense with technical detail, and quite provocative. I don’t like some of the ideas being put forward (LLMs are conscious, for example) but I do look forward to reading Agüera y Arcas justify them.

May 26 | Maintenance: Of Everything by Stewart Brand (4/5)

‘Taking responsibility for something can be a radical act.’ On the surface, Maintenance: Of Everything is about keeping sailing boats and motorbikes in good nick, but there’s more to it — how we maintain ourselves, our relationships and our lives. While I was reading the book, a magnet fell out of the case housing my e-reader (the cheapest one on Amazon). Previously I might have lived without, or bought a new one. Instead, I fixed it. Now I feel a small twinge of pride whenever I use the e-reader.

The book is beautifully designed and put together, and it includes interesting sidebars from authors that commented online as Brand openly drafted the book on Works in Progress. Strong recommend.

Apr 26 | Cat’s Cradle by Kurt Vonnegut (4/5)

Very short, very weird, very funny, a little melancholy, hard to describe. My only regret is that I gave it away, so I’ll have to buy it again for a reread.

Apr 26 | The Scientist in the Crib by Alison Gopnik, Andrew Meltzoff, Patricia Kuhl (3/5)

The Scientist in the Crib is a fairly short read by three Childhood Development researchers. The book describes how young infants learn to piece the world together. The identify three core aspects of childhood learning:

• Representation learning, driven by genetics (e.g. object-centric vision, distinguishing speech patterns)
• Learning via physical and linguistic experimentation
• Interaction with other people — both learning from them (that’s two-way — think about the exaggerated, easily distinguished mother-ese we naturally use with babies), and about them (e.g. mentalisation, Theory of Mind).

They hypothesise that we exhibit a curiosity drive that compels us to explain things we see, that we feel as powerfully as any other biological impulse (take your pick from the Four F’s — fight, flee, feed,… reproduce). One memorable example from the book — if you connect a baby’s foot via a silk ribbon to a mobile, they will learn to kick their foot to make it move. If you remove the ribbon, they’ll try to kick but find it no longer moves. Puzzled, they’ll resort to another trick — one that works on Mum — they’ll earnestly smile at the mobile, to see if that moves it. That’s while they’re still figuring out causality I guess. Either way, a good read, although I wish it had a little more scientific detail.

Apr 26 | On Writing: A Memoir of the Craft by Stephen King (4/5)

A great discussion of writing — think Strunk & White’s classic style guide, only with more depth and examples. It helps that King’s writing is so fun to read — it’s evocative, it’s conversational, it’s intelligent. He starts with a brief memoir that I would have happily read by itself. It’s followed by more specifics on writing, naturally with a focus on fiction (I reckon it carries over to nonfiction too though). The general thrust is to write lots, read lots, and be honest. The best part is the closer, in which he offers an unedited and edited version of the same passage — the second version comes alive. A fairly easy ‘recommend’ for anyone interested in the craft of writing.

Some memorable directions:

• No adverbs: ‘to write adverbs is human, to write he said or she said is divine’
• Active beats passive (‘Frank killed John’ beats ‘John was killed by Frank’)
• ‘You must do two things above all others: read a lot and write a lot. There’s no way around these two things.’
• ‘Reading takes time, and the glass teat takes too much of it.’ Lots of detail on setting up a workspace and limiting distraction so you can work hard.
• ‘What are you going to write about? Anything at all… as long as you tell the truth.’
• ‘The basic rule of vocabulary is always use the first word that comes to your mind, if it is appropriate.’
• ‘2nd draft = 1st draft — 10%’
• Let a piece sit for six weeks after the first draft, and do not touch it or look at it. It will feel like someone else’s piece when you return, and it is easier to kill someone else’s darlings than your own.

Mar 26 | Life Ascending: The Ten Great Inventions of Evolution by Nick Lane

Perhaps I was overly excitable after crashing through A Brief History of Intelligence, but Life Ascending isn’t for me. It’s mostly about the biochemistry of the origin of life, which although interesting, I didn’t find very inspiring — I ended up dropping the book ~150 pages in.

Mar 26 | A Brief History of Intelligence by Max S. Bennett (5/5)

The book breaks down the evolutionary innovations underpinning intelligence. It steps through five innovations, with parallels in AI: steering (RL), pattern recognition (ML), simulation of your own actions (world models), simulation of others’ state (Multi-Agent RL and Theory of Mind), and language (NLP). Quite a few research ideas struck me while reading the book.

As someone working on AI research, this book is a humbling read — the human brain achieves all the things we’d like to do simultaneously, efficiently, in a world with no resets. The more you learn about the brain and the body, the more AI and robotics looks like a poor imitation. Modern AI is improving rapidly, but we have no unified model for all five of Bennett’s innovations, and nor do we seem very close. Cannot recommend strongly enough, for anyone interested in the nature of intelligence. Thanks to Tim Röcktaschel for including a recommendation for this book in his UCL course on Open-Endedness.

Feb 26 | The Art of Doing Science and Engineering by Richard Hamming

In 1995, Turing Award winner Richard Hamming gave an incredible talk titled ‘You and Your Research’. In the talk, he laid out the patterns he had observed which lead to successful research careers — what enabled people to do good work. This book takes the lecture and stretches it out with biographical detail, covering his life’s work. Expect anecdotes about error correction, mainframes, and FORTRAN. Personally I dropped the book after getting about 2/3rds in — I don’t think it’s worth reading unless you’re interested in Hamming’s era of computing. The lecture, however, is essential viewing.

Jan 26 | The Siege by Ben MacIntyre (3/5)

A gift from Dad, not the sort of thing I would pick out for myself. But it’s a good romp about the Iranian hostage situation in London (not to be confused with a similar story at the US embassy in Tehran). The book gives some context on the Iranian Revolution and subsequent downfall of the Shah, which has unfortunately turned out to be quite relevant at the time of writing, but adds some great colour. I preferred the first 2/3rds of the book, which covers the situation within the embassy. As the tension ratchets up, it’s hard not to feel some sympathy for the hostage-takers: as their situation worsens, it becomes increasingly apparent that they’re in far over their heads. The final third details an SAS rescue attempt, characterised mostly by total chaos. A nice page-turner.

Dec 25 | Human Compatible by Stuart Russell (3/5)

A good read by one of the pioneers of Inverse RL, in which an agent tries to infer the reward function a human optimises for. In Russell’s framing, safe, aligned AI will need to truly understand complex human preferences. The risk of pure reward maximisation (as in standard RL) is opening yourself up to reward hacking. Inverse RL comes with its own challenges though: how do you handle human preferences that change constantly, whether due to uncertainty, or just shifts in what we value?

The book is well-written and clearly communicated, but for me it doesn’t fully address the question of capability. RL’s great successes have been agents which learn to exceed human capability. Doesn’t inverse RL cap you out? Think about AlphaGo, DeepMind’s Go-player pretrained on human data, vs. AlphaZero, which was trained with no human data. AlphaZero took longer to train, but its asymptotic performance exceeded AlphaGo’s. How much capability would we be willing to concede for the sake of a more human behaviour? Overall, a good read, would recommend.

Feb 25 | Why Greatness Cannot be Planned by Ken Stanley & Joel Lehman (5/5)

A discussion of invention and discovery, centred on a machine learning experiment named PicBreeder. The authors argue that interesting discoveries don’t happen because of straight shots from A to Z; rather, intermediate results that enable discoveries are unpredictable, forming stepping stones to new ideas. An example from the book: Suppose you’re some God three billion years ago, and you want to develop an intelligent species from the primordial soup. Should you start administering the SAT to single-celled gunk in the hopes of breeding big-brained bacteria? Probably not. To develop intelligent life, many apparently-unrelated steps had to happen first: meat-eating, warm-bloodedness, aerobic respiration, eyes, legs, lungs, and more.

Stanley and Lehman’s point, then, is that we should focus on novel or interesting discoveries which are within reach, rather than aiming for distant goals which we can’t even sketch a route to. In ML terms, they advocate evolutionary methods and novelty search over objective-driven optimisation. Beyond ML, they argue that we should embrace serendipity and permit ourselves to follow our interests. A fantastic book which altered my perspective on decision-making in both ML and real life.

To Write Up

• Oct 25 — The Experience Machine by Andy Clark (4/5)
• Jan 25 — Thinking, Fast and Slow by Daniel Kahneman (3/5)
• Feb 25 — Abundance by Ezra Klein (4/5)