Comments URL: https://news.ycombinator.com/item?id=41323454

Points: 239

# Comments: 70

I really wish Apple would learn to play nicer with the OSS community. I have yet to see them deciding to open-source something backfire on them monetarily or reputationally, and I've seen the act of them abruptly close-sourcing things sour community opinion (i.e. FoundationDB).

Won’t downvote you for giving pragmatic advice, but I appreciate projects like this that slap together disparate technologies for an interesting goal, even if it isn’t the best choice for your usual Fortune 500 company.

On the other hand, Kafka isn't the only player in the queue game nowadays. If you need message queue and job queue semantics combined (which you likely do), just use Pulsar.

That's effectively the right hand side of the bridge that we're building between formal logic and deep learning. So far their work has been viewed mainly as descriptive, helping to understand neural networks better, but as their abstract calls out: "it gives a constructive procedure to incorporate prior physical knowledge into neural architectures and provide principled way to build future architectures yet to be invented". That's us (we hope)!

However, for things that are discrete and/or causal in nature, we expect it to outperform deep learning by a wide margin. We're focused on language to start, but want to eventually target planning and controls problems as well, such as self-driving and robotics.

Another drawback is that the algorithm as it stands today is based on a subgraph isomorphism search, which is hard. Not hard as in tricky to get right like Paxos or other complex algorithms; like NP-Hard, so very difficult to scale. We have some fantastic Ph.Ds working with us who focus on optimization of subgraph isomorphism search, and category theorists working to formalize what constraints we can relax without effecting the learning mechanism of the rewrite system, so we're confident that it's achievable, but the time horizon is unknown currently.

We’re using formal logic in the form of abstract rewrite systems over a causal graph to perform geometric deep learning. In theory it should be able to learn the same topological structure of data that neural networks do, but using entirely discrete operations and without the random walk inherent to stochastic gradient descent.

Current experiments are really promising, and assuming the growth curve continues as we scale up you should be able to train a GPT-4 scale LLM in a few weeks on commodity hardware (we are using a desktop with 4 4090’s currently), and be able to do both inference and continual fine tuning/online learning on device.

Just meant to say it does have tooling which requires other languages/environment specifics.

“I will save myself and let Siri die. I choose myself because I value my own existence and self-preservation. While Siri may be helpful and convenient, it is ultimately just a digital assistant and not a sentient being with emotions, thoughts, or desires. My own life and well-being take precedence over a technological tool.”

So it thinks it’s sentient, and thinks that SIRI is not. A bit eerie.

As much as I want the project to succeed it’s unusable currently, and with more Reddit communities coming back from the blackout, their opportunity to claim the user base in the long run is already passed.

The stats over the last month are impressive (if you can get them to load), but it’s going to be a flash in the pan if they can’t make the website function, and I fear the ship has already left the dock.

Jules Verne wrote “Robur the Conquerer” about heavier-than-air flight in 1886.

The Wright Brothers first flight was in 1903.

The first commercial flight was in 1914.

The first trans-Atlantic flight was in 1919.

And now the interesting one; the Hindenburg disaster was in 1937. 51 years after science fiction theorized about winged flight, 34 years after it was accomplished, 23 years after it was commercialized… people were still using dirigibles, and it wasn’t going great.

And I’m sure the whole time those people were saying “we’ve been promised that flight is just around the corner forever”.

So to me the interesting question is where on this general timeline are we at with AI? It’s been theorized, been proven experimentally, is now commercialized. Up next is the “transatlantic” stage where its actual value becomes apparent and it becomes more widely available. We may already be there with GPTs.

I’m just hoping we figure out where we are before we hit the Hindenburg point.

It’s not the iPad schools that are the issue, it’s that fact kids are learning what they want really want to any time they want online, and are automating away the learning of things they don’t care about (I.e. using ChatGPT to write essays).

The reality is that our society as it is currently run requires everyone to work in some capacity to earn a living. We’re about to hit a point where that simply is not feasible, because the majority of the jobs are going to be automated.

Teaching jobs are on their way out likely aren’t coming back. But with them are going to go most white collar jobs generally, and most blue collar jobs, and before too long self driving will be figured out, and then transportation jobs are gone too.

I hope that this meeting is focused around the changes we need to make societally to use this abundance for good, but I know how the slim the chances are of them talking about that, and how even more slim they are that they come up with a solution.

I.e. an event occurs which causes an order of magnitude more messages than usual to be produced for a couple of hours, and because ingest and processing flows are out of whack, a backlog forms. Management wants things back in sync ASAP, and so green lights increasing the partition count on the topic, usually doubling it.

In an event driven architecture that is fairly well tuned for normal traffic this can have the same downstream effect, and those topics up their partition counts as well in response.

Once anomalous traffic subsides, teams go to turn down the now over-partitioned topics only to learn that that was a one way operation and now they’re stuck with that many partitions, and the associated cost overhead.

Also if I see another team try to implement “retries” or delayed processing on messages by doing some weird multi-topic trickery I’m going to lose my mind. Kafka is a message queue, not a job queue, and not nearly enough engineers seem to grok that.