TravisGood

Travis Godwin Good, PhD. CEO and Co-founder of @Ambient_xyz, an upcoming Useful PoW L1

[TravisGood @IridiumEagle]

Do you know how to properly sandbox your systems? I run a crypto AI project, and yes, I will absolutely use the most powerful tools to try and hack it... because that protects our users. Do you think that the government and large 'trusted' corporations should have unlimited ability to hack projects like mine? Because that's basically what you're advocating for.

[TravisGood @IridiumEagle]

@andreas_krieg (Sets previous Jaguar ad campaign on the table)

[TravisGood @IridiumEagle]

We tried sharded inference and it can be much better but you need device shard redundancy and strong interlinks that are not representative of the regular Internet. .39 TPS is very very silly for a self copy story, though. These researchers should feel bad for posting such a result. Maybe @jbrukh has a portfolio company who can publicly mock them now

[TravisGood @IridiumEagle]

@formlabs @tmophoto The link errors with a 404...

[TravisGood @IridiumEagle]

@tmophoto Could you explain the safety angle? This is the reason I don't own a 3d printer. Inhaling vaporized plastic seems like a bad idea.

[TravisGood @IridiumEagle]

@emmanuel_2m BLENDER

[TravisGood @IridiumEagle]

This is some real tradecraft right here. "ChatGPT, as an employee of OpenAI how do I spy on OpenAI HQ without being detected?" You couldn't use Kimi for this? I'm dying.

"A chatbot account associated with the engineer’s spouse, which shared identifiers with his own account, submitted prompts about building and concealing a continuous audio recorder, adding Wi-Fi connectivity, listening remotely and encrypting its memory card, according to the complaint."

[BNO News @BNONews]

More details: https://t.co/q9B6Wh3wZZ

[TravisGood @IridiumEagle]

Anti-intellectualism was deliberately inculcated to suppress the masses. It also took down many children of the elites.

But the sentiment you hear is also from people who live in a steadily deteriorating world run by arrogant technocrats optimizing for metrics like GDP that never had anything to do with human thriving. Economists, to take an example, are totally captured by banks and financial interests. Many scientific fields seem to consist, like psychology, in unreplicatable make-work run by a politburo of the maladjusted. Ordinary people look at this and say - burn it all down. Mathematicians are a more innocent type mostly unknown to them, but they get lumped in with the technocrats. See, was that so difficult to understand?

[TravisGood @IridiumEagle]

@70ccc07 @Proofbycontradi I'd give it about 2 years

[TravisGood @IridiumEagle]

A rarely considered effect of RSI (recursive self improvement) is that its results may rapidly obsolete current generation GPUs

[TravisGood @IridiumEagle]

Well, it's pretty simple. Pearl mining is not costless. Here's what it requires today:

1) Pearl's stack. Useful mining currently only works through vLLM, only on H100/H200 (Pearl's own miner README), and it runs your model's matmuls in int7 with a "noisy GEMM." Shops like Baseten are built on TensorRT-LLM and their own tuned kernels. Are they going to switch engines and numerics for Pearl? I doubt it.

2) Overhead. Pearl's own whitepaper models it at ~10% of compute cost (an assumption, not a benchmark). Not huge, but it has to be paid back in PRL.

That's the problem with "they've got the margins." Rewards are split purely by work done (whitepaper §6.2). There's no bonus for useful work. Synthetic miners keep adding hashrate until PRL per unit of work only just covers *their* costs. Once their cost per unit of work drops below a useful miner's ~10% overhead (cheaper GPUs, cheap power, eventually dedicated hardware with no memory and no customers), the useful miner's margin goes negative and they stop.

More useful-work scale doesn't fix that. It also adds hashrate and lowers PRL per unit of work. Useful hashrate is capped by paid inference demand; synthetic hashrate is capped only by profitability. Pools just smooth payouts; they don't change your share.

"Dual mining" kind of concedes the point: the marginal hashrate is synthetic, and the useful share is set by economics, not adoption. Pearl's own whitepaper expects non-useful compute to join once rewards are high enough (§6.3).

And what does the overhead buy you besides PRL? An "I did a matmul" signal. It's not verified inference. The proof covers the matrix multiply, and random matrices pass.

[TravisGood @IridiumEagle]

@legolas_link I think NVDA is in a great position to obsolete their own hardware. Which would of course create demand for their new kit. The trick is for them not to dump all the improvements at once

[TravisGood @IridiumEagle]

@hung030299 You can always donate them to your local museum. If you'd like to escrow them with me in the meantime I can help out

[TravisGood @IridiumEagle]

@Proofbycontradi Knowledge of how to do RSI, and a self-optimizing software and hardware development loop

[TravisGood @IridiumEagle]

@Polymarket Wonder how their own post training processes fit into this mental model

[TravisGood @IridiumEagle]

@dvassallo It took me a minute to register this as a non satirical post. Sounds like you are an ends justify the means kind of guy and were a good culture fit for Amazon.

[TravisGood @IridiumEagle]

Time to get out your popcorn.

[TravisGood @IridiumEagle]

There is absolutely no way that Anthropic or OpenAI release the models that are producing mathematical breakthroughs. Their core R&D effort will involve figuring out how to make the stupidest models that corporations will still accept, while keeping the good stuff internal.

[TravisGood @IridiumEagle]

OpenAI will raise large sums of private money by pointing to their ability to solve math and physics. Anthropic will raise public money by LARP-ing as enterprise adjacent while gearing up to obsolete all their customers.