Frontier.

Then

A processor of the 8086 generation is a single clerk at a single counter. A program is the queue: instructions arrive one at a time, each is fetched, decoded, executed, and the next one waits. Speed is a matter of fewer and cheaper instructions, and the assembly programmer's craft is knowing the price of every one: which register holds what, how many cycles a multiply costs, when a memory access will stall the clerk. The Intel 8086 of 1978 had about 29,000 transistors and ran at 5 MHz; forty years of Moore's law made the clerk astonishingly fast, but it stayed one clerk. This is latency-optimised computing: do one thing, then the next, as quickly as possible.

Now

A modern accelerator opens tens of thousands of counters at once and asks a different question: not "how fast is one instruction?" but "how do I keep all of these fed?" Nvidia's model, SIMT (single instruction, multiple threads), runs the same small program over enormous arrays of data in lock-step; the unit of optimisation is no longer the instruction but the kernel. The Blackwell chip announced in March 2024 carries about 208 billion transistors on two dies and moves several terabytes of data per second in and out of memory Nvidia, 2024 ↗. And that is the tell: the binding constraint is not arithmetic any more but memory bandwidth, the corridor between the counters and the warehouse. This is throughput-optimised computing.

What actually connects them

The question is unchanged: how do you organise computation against a scarce physical resource? The answer inverted, control flow gave way to data flow and latency to throughput, but the discipline that answers it is the same. A fast CUDA kernel and a slow one differ in exactly the way a good and a bad assembly routine differed: whether the author knows what the hardware actually does with the bytes. Memory layout, cache lines, the cost of a branch. This is a genuine lineage of craft, not of architecture. The GPU is not the descendant of the 8086; the programmer who understands both is.

Where it crosses

• → Learning: the GPU is the reason a 1980s algorithm became a 2020s industry (02).
• → Quantum: quantum systems are designed, simulated and controlled with classical accelerated compute (04).
• → Trust: proof-of-work mining drove a generation of GPU and ASIC demand (05).

Then

Two ideas came out of the same 1980s moment, and this section follows both. The first is an idea: in 1982 John Hopfield showed that a network of simple, symmetrically connected units behaves like a physical system rolling downhill in an energy landscape, and that its valleys can be made to store memories, so that a partial or noisy pattern relaxes into the complete one Hopfield, PNAS 1982 ↗. Content-addressable memory as physics. Four years later, backpropagation gave multi-layer networks a way to learn Rumelhart, Hinton & Williams, Nature 1986 ↗. The second is a substrate: Carver Mead's analog VLSI programme at Caltech argued that if computation is physics, the transistor's own physics should do it: subthreshold currents as exponentials, Kirchhoff's law as addition, a silicon retina that senses and processes on one chip, in real time, at milliwatts Mead, Proc. IEEE 1990 ↗. Both were right. Neither had the compute, or the data, to prove it at scale.

Now

The idea won, and it won on GPUs. AlexNet in 2012 trained a deep convolutional network on two consumer graphics cards and cut the ImageNet error rate nearly in half; the transformer arrived in 2017 Vaswani et al. ↗; large language models followed. And the continuity is unusually literal: the attention mechanism at the heart of every transformer is mathematically the update rule of a modern Hopfield network with continuous states and exponential storage capacity Ramsauer et al., 2020 ↗. In October 2024 the Nobel Prize in Physics went to Hopfield and Hinton for the foundations laid in those years Nobel Foundation ↗. Meanwhile the substrate is coming back for a reason the idea created: energy. Data centres drew roughly 1.5% of global electricity in 2024, AI the fastest-growing share IEA, 2025 ↗. Neuromorphic and analog in-memory chips (Intel's Loihi 2 and the 1.15-billion-neuron Hala Point system (2024), IBM's NorthPole Science, 2023 ↗, Manchester's SpiNNaker) are Mead's argument returning as engineering.

What actually connects them

This is the one of the five with a loop in it. The idea and the substrate left the same room in the 1980s and went separate ways: the idea lay dormant until digital hardware, the GPU, could feed it; the substrate lost to that same digital hardware and was largely abandoned. Then the idea grew so large that its energy bill became the argument for the substrate that had lost. Strand A won; strand A's cost is reviving strand B. Two honest caveats. First, "attention is a Hopfield update" is an exact mathematical statement about a modern generalisation, not a claim that transformers were derived from Hopfield's 1982 paper; the lineage is real, but it was recognised in hindsight. Second, neuromorphic hardware has been "returning" for a decade; the energy wall makes the case stronger than ever, but the substrate has not yet won a mainstream workload. Watch that, not the press releases.

Where it crosses

• ← Compute: the hardware that unlocked the idea, and the energy wall it produced (01).
• → Brain–machine: the decoders behind today's implants are deep networks (03).
• → Trust: verifiable inference and data provenance are where AI's credibility problem meets the ledger (05).

Then

In 1991 Peter Fromherz's group laid a single leech neuron across the gate of a field-effect transistor and recorded its action potential through the chip Fromherz et al., Science 1991 ↗. Everything hard about the field was already visible in that experiment. The cell does not touch the silicon: a cleft of electrolyte, tens of nanometres wide, sits between membrane and oxide, and the neuron's signal reaches the transistor only as the faint extracellular voltage that its ion currents drop across that gap: microvolts, through an impedance nobody controls, drowned in noise. Coupling can be capacitive (clean, weak) or faradaic (stronger, but chemistry happens at the electrode). And a living cell treats a foreign surface as an injury: it walls it off. Signal-to-noise, biocompatibility, chronic stability: the vocabulary was fixed in the first decade.

Now

Thirty years on, the numbers have moved by orders of magnitude. The 96-electrode Utah array that gave a paralysed man cursor control in 2006 Hochberg et al., Nature ↗ has become Neuralink's N1: 1,024 electrodes on 64 flexible polymer threads, placed by a surgical robot, transmitting wirelessly, first implanted in a human in January 2024 Neuralink ↗. Competing routes avoid the brain surface altogether: Synchron's stentrode rides in a blood vessel; Precision Neuroscience's film sits on the cortex without penetrating it. And the decoders are deep networks: in 2023 two groups restored speech at 60–80 words per minute from cortical recordings Willett et al. ↗, Metzger et al., Nature ↗.

What actually connects them

The interface physics has not changed. What scaled is electrode count, materials, surgery and, above all, the decoder, which is now a deep-learning problem borrowed wholesale from Learning (02). What did not scale is the wet gap. The foreign-body response that walls off an electrode over months is the same biology Fromherz's neuron faced; Neuralink's own first patient saw a majority of threads retract from the tissue within weeks of implantation, recovered by software Neuralink, May 2024 ↗. This is the most sobering section on the page and the honest counterweight to the hype: the frontier is real, the demonstrations are extraordinary, and the hard problem is still the one the field started with: a chronically stable interface between a living cell and a dead one.

Where it crosses

• ← Learning: the decoding stack is deep learning; and neuroscience keeps supplying the metaphors that AI then over-claims (02).
• ← Compute: on-device inference at implant power budgets is Compute's problem in miniature (01).

Then

Grow a crystal one atomic layer at a time, gallium arsenide and then aluminium gallium arsenide, and electrons collect in a sheet at the boundary so clean they travel micrometres without scattering: a two-dimensional electron gas. Put a metal gate above it and a voltage shapes where those electrons may go; cool it to millikelvins and add a magnetic field, and their quantum nature is laid bare. This is the physics that produced the quantum Hall effect (1980) and its fractional cousin (1982), two Nobel Prizes, and a generation of experiments in which the gate was an instrument, a way of asking electrons what they do. Our own questions: what happens when electrons move through a lattice we built to order, a superlattice? They show Bloch oscillations, minigaps, magnetic breakdown, and interference between paths, as if the crystal were an interferometer. And, with an electron gas tuned into the fractional quantum Hall regime, the tiny hyperfine coupling between electron spins and the nuclear spins of the host atoms could be switched on and off with a gate voltage.

Now

The same gate-defined electron gas is now a fabrication method. In 1998 Loss and DiVincenzo proposed holding one electron in an electrostatically defined dot and using its spin as a qubit PRA 1998 ↗; by 2022 silicon spin qubits had crossed 99% two-qubit fidelity Xue et al., Nature ↗, and Intel is fabricating them on 300-mm wafers. The larger machines today are superconducting: in December 2024 Google's 105-qubit Willow processor ran a surface code whose logical error rate fell as the code grew: error correction below threshold, the condition every fault-tolerant roadmap depends on Nature 2024 ↗. A third route, topological qubits built from the non-abelian excitations predicted at filling factor 5/2 in the fractional quantum Hall regime, remains the most elegant and the least demonstrated; claims of progress there are contested and should be read as such.

What actually connects them

The tightest continuity of the five. The techniques did not merely inspire the frontier; they are the frontier's toolkit. Electrostatic gating of a two-dimensional electron gas, once the way to study electrons, is the way spin qubits are made. The electron–nuclear hyperfine coupling that was a thing to measure then is the dominant decoherence channel now, which is why the best spin qubits sit in silicon isotopically purified to spin-free silicon-28, engineering away exactly the interaction that made the 2002 experiment interesting. And the fractional quantum Hall states studied at millikelvins are the theoretical root of the topological programme. Instrument, product: same physics, different purpose. The honest limit is on the other side: none of this tells you when a quantum computer will do useful work. Below-threshold error correction is a necessary condition, not a schedule.

Where it crosses

• ← Compute: qubits are designed, simulated and controlled with classical accelerated compute (01).
• → Learning: "quantum machine learning" is a research programme, not an established advantage, flagged as speculative (02).

Then

The problem is old and precise: how do parties who do not trust each other agree on a shared record? Byzantine fault tolerance had answers for small, known groups; Bitcoin's 2008 white paper answered it for anonymous strangers by making lying expensive, proof of work, and chaining hashed blocks so that rewriting history costs more than it earns Nakamoto, 2008 ↗. Ethereum (2015) generalised the ledger into a machine that executes code, and by 2017 tokens were being used as coordination mechanisms for whole markets, DENT among them. The constraint was equally precise: the trilemma. Scale, security, decentralisation: pick two. Every chain paid for trustlessness with throughput.

Now

The ambition of the "world computer", everything on-chain, ran into the trilemma and has been re-engineering around it since. Rollups execute off-chain and post compressed results; zero-knowledge proofs let a party prove that a computation was done correctly without re-running it or revealing its inputs: validity proofs small enough to check on a phone for work that took a data centre. That is the durable idea underneath the noise: verifiable computation without a trusted party. It is now meeting problems that need it, and several are AI's: proving that a specific model produced a specific answer, attributing training data, running inference on hardware you don't own and trusting the result. Where the technology has found unambiguous product–market fit is more mundane: dollar stablecoins, settling trillions of dollars a year on-chain by most counts and given a US federal framework in July 2025.

What actually connects them

Say it plainly: this is the weakest lineage of the five and the most contested endpoint. Blockchain did not produce today's frontier the way transport physics produced qubits. What it did was pose a question, computation you can verify without trusting the computer, well before there were problems worth the cost of answering it. The ambition outran the substrate; most of what was built on the substrate in 2017 did not survive. The question survived, and its best answers (zero-knowledge proofs, in particular) are only now becoming cheap enough to matter. The honesty here is what makes the other four credible.

Where it crosses

• ← Compute: mining economics drove GPU and ASIC demand for a decade (01).
• → Learning: verifiable inference and provenance are where the ledger meets AI's credibility problem (02).