Josh Syer

[Josh Syer @SyerJosh]

@chrislevan24 please release this yesterday

[Josh Syer @SyerJosh]

@tokifyi at ubc

[james hawkins @james406]

introducing my new benchmark, "AI models randomly sorted on an X and Y axis"

[Josh Syer @SyerJosh]

@JustinasRoland0 Just added Claude token interest rates to my watchlist

[Josh Syer @SyerJosh]

@imnotchalk not very pacing the frontier of you

[Pranav Vinod @pranav_vinodrg]

The electrons must flow.

I have spent the past few months mapping the energy to compute space, following the whole trip an electron takes, from the moment it leaves a generator to the moment it becomes a token. It started as a side project and slowly became most of what I read. The more I understood it, the less it looked like a set of separate markets and the more it looked like a single flow.

An electron leaves a turbine at about 18,000 volts and is pushed up to 345,000, because high voltage is how power crosses distance without bleeding away in the wire. It runs a hundred miles of line, drops through a substation, and passes through a transformer the size of a school bus into the building.

Inside, it falls back down. To 800 volts, then to 50, then below one, until it reaches the single volt a chip runs on. Some of that stepping is unavoidable, but a lot of it is not. The grid runs on alternating current, built a century ago to push power to motors. The chip wants direct current. So the standard path flips one into the other and back more than once. Every flip runs through a switch that is never perfectly on or off, and every one loses a little energy as heat. A good chain from the fence to the chip still arrives below 90 percent. Some of that loss is physics. Some is just history, and history can be deleted.

The electron then reaches the chip, and most of the power it draws goes to waiting, not computing. To write a single token the chip reads billions of numbers out of memory and does a little math on each. Reading those numbers takes far longer than doing the math on them. A top chip finishes the math for a token in about fifteen microseconds. Then it waits four and a half milliseconds for the next numbers to arrive, drawing nearly full power the whole time. No frontier model fits on a single chip either, so it is split across many, each one reading from its own memory and reaching across to the rest. The gap is not an accident. Over the last twenty years the math a chip can do grew about 60,000x, while the speed of feeding it data from memory grew only 100x, and the links between those chips only 30x. Doing the math got cheap years ago. Feeding the chip the data it needs became the hard part. Even on a carefully tuned run the best teams in the world use only 38 to 43 percent of the chip they are paying for. So after all that work of carrying a clean watt to the silicon, the silicon spends most of its life waiting. Whatever it does produce leaves the building as a token, as light down a fiber.

Seen whole, the same two problems run its entire length. It’s a lossy inefficient flow that leaks energy and performance, and it waits in a queue at almost every stage. The median power plant that switched on last year waited 61 months just to connect to the grid. A substation transformer takes over two years to arrive. Gas turbine slots are booked out to 2031. Even the chip waits, idle about a fifth of the hours it is powered on.

Once I saw it this way, the companies worth watching all did one of two things. They shorten the flow, or they own a piece of it. Remove a loss, and more of each watt reaches the chip instead of leaking away as heat, which means more tokens per joule. Remove a step, and you skip both its energy loss and the wait for the equipment it needed. Remove a wait, and you pull a project's whole revenue forward, which is worth the most right now. A single year of a gigawatt pulled forward is worth over two billion dollars. That is why the same megawatt-hour sells for $271 to a buyer who can wait and nearly $5,000 to one who cannot. Owning is different. You hold a point the flow cannot go around, built of physics or law, and you charge whatever the detour would cost. The best companies do both, in order. Shorten to enter, own to stay.

That frame is what the whole map and the essay are built on. About 450 startups so far, mostly North American, sorted across the eleven stages of the flow. It is not the full list, and it keeps growing.

If you are building anywhere on this flow, I would love to hear from you!!

The full essay and the map are in the comments.

[Josh Syer @SyerJosh]

@ellaaborating And there’s more to come👀

[Josh Syer @SyerJosh]

In 2025 and 2026, at least seven companies with UBC founders got into Y Combinator. Each one a potential unicorn.

What if the next one could start here?

The UBC Tech Map goes live Tuesday 13 October.

[chris @chrislevan24]

if you can’t be bothered to read here’s a 60 second summary.

[Josh Syer @SyerJosh]

@chrislevan24 @frontrowvc FOMO

[Alex @AlexanderTw33ts]

Today we're launching https://t.co/wOE9a8V4gN

Jev is our latest model, state of the art in zero-shot learning and real-time reasoning. For the next 10 hours, Jev is available live at 100% uptime. $5 per query.

[Pranav Vinod @pranav_vinodrg]

gang gang

[Josh Syer @SyerJosh]

See you there!

[Pranav Vinod @pranav_vinodrg]

Tomorrow at 9, my friends @frontrowvc are running with @_yasharya1 and Founders Running Club Vancouver.
​
If you’re a founder, operator, builder or anyone who just wants to run with some cool folks pu!!!

Classic Vancouver 5k run across the English Bay seawall and end at The Grind by VEGAIN.

☀️☀️🏃🏃‍♀️

[Josh Syer @SyerJosh]

@tokifyi @pranav_vinodrg @frontrowvc @_yasharya1 Is agi at the end of the run?

[Josh Syer @SyerJosh]

@pranav_vinodrg @buildcanada Question is: will we become a compute superpower with our NG, the third largest LNG exporter, or both

[Mateo Q. @Mateoqr_]

Good times with the @frontrowvc squad last weekend!

[Josh Syer @SyerJosh]

@chrislevan24 @frontrowvc 5 months in and joining has been the best decision I have made in university

[Dario Amodei @DarioAmodei]

We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so.

Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models’ alignment during training.

You can read the full post here: https://t.co/OGyPb7yaYt

[Josh Syer @SyerJosh]

@tokifyi Someone said it