lettergram

Unfortunately, the AI being locked down and proprietary is the winning strategy for these companies.

My company hosts its own models. Some customers require us to use either US / EU models, while others are fine with us using any model.

As such, we have two GPU clusters, the general AI cluster runs a Chinese model as it's the most accurate and robust. The US/EU required ones have a few percentage points lower on our accuracy metrics and we provide them those that require it for an extra fee.

Why host at all? Because it enables us to get much higher margins than competitors, while reducing costs. Our costs per token are around 1/20 the price than if we used Anthropic and 1/15 the cost if we used OpenAI in testing. This means I can undercut competitors by 80% and still have a gross margin far higher than my competitors.

In reality, these US AI providers are jacking up the prices and trying to implement regulatory capture. I'm actually fairly confident they'll succeed. At some point, I'm expecting the US / EU administration(s) to block foreign based model, at the same time, they'll probably invest in Anthropic and OpenAI.

What Anthropic and OpenAI are doing is using "safety" as a wedge, just like large corporations used "environmentalism" or "food safety" or "workers safety" as a wedge to regulate smaller competitors out of the picture. Then they jack up rates, sue and/or buy anyone who can potentially be a threat. It's the #1 threat to our business model.

Our competitors are giving half of their margin over to these large AI service providers, we keep the vast majority of ours. Eventually the AI service provider will be able to squeeze them even more until the margin just isn't there and either they are purchased or replaced via internal tools at the company they sell to.

If you're self-hosting a model as a startup (e.g. using GPUs), you're almost certainly using a Chinese model. If your a startup outsourcing (e.g. using tokens), you're going to be using a US based model

As a tractor owner. Two things, the DPF & SCR (>=75hp) on a tractor is not a great idea --

1) Tractors are typically owned by low margin businesses (i.e. farmers) that need to be repaired in the field AND need to be repaired quickly, else you loose a crop. Adding complexity to tractors literally can cost the farm.

2) The actual emissions reduced is questionable. Tractors run significantly less than a truck, like 50-100x less often. Further there are at least 2x more trucks sold per year

3) To run the SCR system, the engine had to run hot for like 20 minutes burning extra fuel and required DEF (yet more input costs)

3) The emissions they are trying to reduce with the these are likely not excessively harmful from a tractor; largely because most tractors who need an SCR system is >75hp, which also means they're typically used on a large farm (100+ acres). Which dissipates the risks substantially.

For reference my 2022 Kubota tractor repeatedly had issues with the DPF / SCR system, mostly the software to enforce environmental rules. This lost us ~$20k one year due to the tractor being knocked out for a week (I was mid-cut for 140 acre hay, rained & rotted in the field post-cut).

For reference, I was very much ready to bypass the SCR system, but decided against it to keep the warranty. It had nothing to do about "right to repair", I figured out exactly how to bypass it.

The stats come from the government agency in the US which tracks driver fatalities.

Here's the source document: https://crashstats.nhtsa.dot.gov/Api/Public/ViewPublication/...

Regarding "distracted driving" causing cyclist fatalities, my point was more that bicyclists have a higher risk rate. If we're already going down this path for safety, banning bicycling would have a higher impact than reducing fatalities ~10% which is what the EU claimed it could prevent.

Regarding the claim that "pretty much all fatalities on bikes are the result of being killed by drivers", that's simply not true. It also largely depends on laws and location.

https://www.bikeattorney.com/bike-accident-common-causes.htm...

https://www.npr.org/sections/health-shots/2011/05/20/1364622...

incredible amount of killing from distracted drivers

1.10 fatalities per 100 million vehicle miles traveled

https://www.transportation.gov/briefing-room/trumps-transpor...

For reference, fatalities on bikes:

9.32 fatalities per 100 million cyclist miles (6 deaths per 100M kilometers)

https://www.calbike.org/urban-transportation-research-bike-f...

We actually found the Mistral Small 4, quantized to 4bit was comparable to Qwen 3.6 27B and is roughly the same size. At least from our experience on our use cases, the quantization of the Mistral model worked far better than trying to quantize the Qwen family.

Fully agree to your point though, Mistral in general is far behind where I'd expect and Qwen in particular is crushing it at the smaller sizes.

Personally, I'd consider anything 20B params and above a "medium" model. Small being <20B and large >100B. I think obviously we can get to the huge 1-2T param models, but frankly the margin of accuracy improvement for the speed hit is kinda insane (1-2% for many metrics).

Yup, that's pretty much what I was saying for the above. I had a Kubota and can indeed attest to the issue(s).

They're both considered medium tractors with similar weights for a similar purpose. I forget the official weights, but it's something like 100-175HP is mid-tier.

You can get a kubota M5-111 with a closed cab for $70k-100k, cheaper than these. Plus zero percent financing though then for 5-7 years. Well built and a comparable class in terms of weight and horse power.

People aren’t buying them for price, but the first sentence discusses it as if it’s relevant.

My assumption are farmers are trying to skirt the eco rules for vehicles in some way. Which by the way is insanely annoying and has caused issues for all the farmers I know at one point or another. Worse, you can’t fix the ecosystems on your own so you have to get them serviced costing quite a bit and importantly putting your tractor out of commission for a while. It’s why older tractors have a premium

Clicking the link splits off the ".", which is interesting but necessary.

You can buy books on how to make and obtain chemicals on your own.

Hell here's an Internet Archive book on making explosives

https://archive.org/details/saxon-kurt.-fireworks-explosives....

If you ever chat with older folks pre-90's much of this information was accessible fairly easily. It only changed with the push by the government to crackdown on Waco, Oklahoma City bombing, militias and other related groups. There was then a campaign to make it "normal" to limit free speech on the subjects, where as these books were available before.

I think the whole thing where AI should make information less available is a difficult battle and one which I personally oppose, but do understand. Free speech and information isn't the problem, it's the people, actions and substances they create.

After the age of the internet, I think it's been a forever loosing battle to limit information, it's why we couldn't stop cryptography, nuclear weapon proliferation, gun distribution, drug distribution, etc. The AI is just another battle ground, one which, if they actually do manage to control could definitely create some walls to this information, but not stop it.

More scary, is that the AI as it becomes pervasive and stop people from asking certain questions, because they don't know they should ask... but that's unrelated to the risk of mass death.

There's still plenty of moats frankly, same moats as before. What isn't the moat is the software development time.

In our case, we're building a tool that has a moat from: integrations, multiple parties connecting, and others

It's very sticky once we get in, and has nothing to do with the software so much as legal, company policy and inter party communication

When I read that I'm always personally confused. He had a commanding voice and had an aurora of being above it all. But when you listened and watched what he actually did, he seemed very political in my mind, though perhaps more of a moderate(?).

He even advocated for world government, endorsed politicians, etc.

1000 downloads lol

https://play.google.com/store/apps/details?id=com.freedomcha...

All 1000 downloads...

https://play.google.com/store/apps/details?id=com.freedomcha...

Feels a little like clickbait "MAGA-themed", never heard of Converso.

That said, the analysis itself is interesting and worth a look, if nothing else it's a general pattern you can follow for many chat applications to see how secure it is.

Bourdain actually joked about killing himself in the exact manner and location in, which he did. When I heard it happened, my wife and I both recalled the same times he'd mentioned it. It wasn't a surprise really.

Bourdain had been referencing Hunter S Thompson and the way he went out for years. He'd also repeatedly mentioned wanting to go out in southern France after a great day. Bourdain generally had the same "vibe" as Thompson as well. Here's Thompson's last note to his wife:

No More Games. No More Bombs. No More Walking. No More Fun. No More Swimming. 67. That is 17 years past 50. 17 more than I needed or wanted. Boring. I am always bitchy. No Fun—for anybody. 67. You are getting Greedy. Act your old age. Relax — This won't hurt.

To me, it wasn't a surprise at all. My wife and I even had discussed when we thought it would happen. The main thing about Bourdain was that people could relate to him and he wrote excellent prose. He seemed authentic and he went out on his terms, which is what he wanted and was the way he lived.

There’s a lot of indications that we’re currently brute forcing these models. There’s honestly not a reason they have to be 1T parameters and cost an insane amount to train and run on inference.

What we’re going to see is as energy becomes a problem; they’ll simply shift to more effective and efficient architectures on both physical hardware and model design. I suspect they can also simply charge more for the service, which reduces usage for senseless applications.

The irony -- so not only was their system hacked ("hosted onsite"), but then it was also burned down onsite with no backups.

In other words.. there was no point in the extra security of being onsite AND the risks of being onsite single failure point destroyed any evidence.

Pretty much what I'd expect tbh, but no remote backup is insane.

Back in 2016 - 2018 my work at Capital One resulted in a modified C-RNN style architecture that was producing gpt-2 level results. Using that model we were able to build a general purpose system that could generate data for any dataset (with minimal training, from scratch):

https://medium.com/capital-one-tech/why-you-dont-necessarily...

At the time it was clear to all on the team that RNNs, just like transformers later on, are general purpose frameworks that really only require more data and size to function. In the 2018-2020 era and probably today, they are slower to train. They also are less prone to certain pitfalls, but overall had the same characteristics.

In the 2019-2020 I was convinced that transformers would give way to better architecture. The RNNs in particular trained faster and required less data, particularly when combined with several architectural components I won’t get into. I believe that’s still true today, though I haven’t worked on it in the last 2-3 years.

That said, transformers “won” because they are better overall building blocks and don’t require the nuances of RNNs. Combined with the compute optimizations that are now present I don’t see that changing in the near term. Folks are even working to convert transformers to RNNs:

https://medium.com/@techsachin/supra-technique-for-linearizi...

There are also RNN based models beating Qwen 3 8B in certain benchmarks

https://www.rwkv.com/

I suspect over time the other methods my team explored and other types of networks and nodes will continue to expand beyond transformers for state of the art LLMs

They tried it last administration too -- https://en.wikipedia.org/wiki/Disinformation_Governance_Boar...

To be honest, it's been going on for effectively forever.

See operation mockingbird -- https://en.wikipedia.org/wiki/Operation_Mockingbird

One crazy thing you can do with color spaces is dramatically improve detection algorithms and store massive data.

https://austingwalters.com/chromatags/

Think of it this way, a QR code is binary. If you modify color spaces correctly you can get 6 bits (or more) per pixel. In addition, you can improve the detection at distance, localization for robots, and speed (120 fps).

Done this to great effect previously and you can do a lot of awesome things with it. Pretty much the easiest hack in computer vision.

Also true, kind of ment the chip sanctions. They transitioned to China and domestic production.

Yes. Strategically, it makes sense for the US, much like Russia and China to be independent.

Sanctions weren’t effective on Russia because they had most of what they needed domestically and partner markets to sell those goods to.

When the US tried to impose sanctions on China, China called the bluff and blocked strategic materials. The US “trade deal” wasn’t much different than how it started.

In terms of willing to pay for it; what’s having a country worth? Because if a competing country can withhold resources you need, you’re effectively a junior partner.

Ultimately, reduce over seas benefits, tariff and offer tax write offs to build on shore. Then you’ll have better higher paying jobs and onshore manufacturing. More real GDP from goods will not have a negative impact or cost, it’s part of why Germany and Japan grew rapidly (they had tight import controls, to build a domestic industry).

Also, the majority of the country voted for Trump and this was his #1 issue. Like him or hate him, the desire for domestic protection is what elected him.

I run a company in this space...

First, China patents ~5-10x more than the US does currently on a given month. Further, China has made it required for companies to patent.

The US definitely could not respect the Chinese patents, or they could treat Chinese patent's differently. IMO there's a ~1% chance of that happening. Patent law is pretty well defined, there are a multitude of treaties and if the US wants their patents to be respected, they have to respect the worlds.

That said, I will say, I suspect a lot of these patents can be invalidated. My company works heavily in this space and we work with some of the top US law firms. We sell a service that's used to identify prior art and invalidate patents in ~15 minutes -- https://search.ipcopilot.ai/

There's a lot of prior art in the open source community that can be used to attack these patents. Further, if folks publish their innovation it'll provide a solid layer of prior art.

For what it's worth, I run a company in the space --

I 100% agree with you and luckily I think with AI this will rapidly change. The USPTO is bringing on as many AI tools as possible, as fast as they can. Similarly, we've built a product that can invalidate patents at scale, conduct prior art searches in 15 minutes what used to take weeks and thousands of dollars --

https://search.ipcopilot.ai/

We and others in the space are rapidly gaining traction, so I suspect it's only a matter of time. I should also mention there are whole networks out there battling patent trolls (LOT Network) and others working on open source, etc.

There’s a reason rural folks have a higher fatality rate. That said, at least in the US, there’s the presumption that those who live more rural are more rugged, capable, and harder working.

I used to live in Chicago and SF. I’ve since moved to rural Tennessee. I can tell you everyone, including my kids, now have learned to drive our tractor. Granted I’m with them, but we had my 4-5 year old moving hay and they were helping me change oil.

I understand the concern, but everyone learns through doing. There’s definitely danger in that, and you should try to limit risk. At the same time; not teaching them is also high risk in that environment, as they’ll do it anyway with friends later.

Agreed here, the description under the headline is important:

A new Danish study finds no association between aluminum in childhood vaccines and 50 different health conditions, including autism, asthma, and autoimmune diseases. The findings reaffirm the safety of Denmark’s childhood vaccination program.

Breaking it down --

1. The study was Denmark specific 2. The study inspected aluminum specifically 3. (Based on reading, it seems to) Reaffirm the findings that the vaccines are safe, conducted / funded by those who set the policy

Here's the actual study: https://www.acpjournals.org/doi/10.7326/ANNALS-25-00997

Note the main researcher is funded by Novo Nordisk: https://researchleaderprogramme.com/recipients/anders-hviid/

And is frequently trying to debunk criticisms of drugs, often with phrasing such as this.

https://www.contagionlive.com/view/hpv-vaccination-does-not-...

https://www.medscape.com/viewarticle/740704

and the list goes on...

The study only covers aluminum:

no association between aluminum in childhood vaccines and 50 different health conditions

It's right in the title, which was omitted on submission.

The concerns the public have are not aluminum, so of course it wont. Should also point out the argument largely made is that the _quantity_ of vaccines, particularly in the US are part of the problem.

Generally speaking, I'd welcome studies on what folks are actually saying. Having friends in both the avid vaccine advocates and the anti-vaccine crowd, I think both are just arguing past each other and not studying stuff rigorously enough to convince the other.

Contrary to this post, I think the AI agents, particularly the online interface of OpenAI's Codex to be a massive help.

One example, I had a PR up that was being reviewed by a colleague. I was driving home from vacation when I saw the 3-4 comments come in. I read them when we stopped for gas, went to OpenAI / codex on my phone, dictated what I needed and made it PR to my branch. Then got back on the road & PR'd it. My colleague saw the PR, agreed and merged it in.

I think of it as having a ton of interns, the AI is about the same quality. It can help to have them, but they often get stuck, need guidance, etc. If you treat the AI like an intern and explain what you need it can often produce good results; just be prepared to fallback to coding quickly.