Cool. In July, ESRI and Google teamed up to lock down road closure data on their “community maps” program. The community appears to be governments feeding their data to Google and ESRI who then look after it for us, entirely for our benefit.
I really think there needs to be an effort to get drivers to use real community software and provide this data as a public good.
I think companies of all sizes will want their own models, or at least customised ones, for their own specific use cases or competition and security issues.
1. Both training and optimisation will get significantly cheaper and easier quickly.
2. Politics will probably get even more insane before a potential reprieve on the 20th of Jan 2029.
3. The big AI firms will become part of the surveillance capitalism network, if they're not already.
So I think for self-protection a lot of companies will be looking near to medium term AI independence.
The argument is sound, but the maths don't math for now, and it's unclear when/if they will.
For the time being, unless you truly have millions, the outcome from training will be very net negative, while focusing on building on top of existing AI will yield amazing things if you apply the same talent and effort.
When it does get cheaper, then it will be easier to acquire the skills and experience too, and the struggle you went through by trying to do it now will be somewhat wasted.
Besides, I am well versed in this field, and it is not rocket science. There are plenty of software engineering domains that are a lot more challenging, like high-end graphics, large-scale data engineering or kernel programming. People will learn to train LLMs when people want them to.
In reality, enterprises are happy to offload even risky tasks to others as long as they get some contractual guarantees about their data. Would they like more choice in who to buy from? Yes, but not enough to in-house such a specific discipline.
The cost of training a model from scratch is going to be cost prohibitive for the vast majority of companies (even if renting the hardware needed for the 1-2 month training time). It's an interesting learning exercise, and some of the things learned can be applied to other parts of the process. There's also the issue of needing a huge amount of data needed to get decent weights.
Fine-tuning a model or LoRA based on the companies data set is more feasible but you're likely going to need several runs as you test/try out different base models, parameters, etc. This is why there are a lot of fine-tuned models on huggingface based on base or instruction-trained models from the larger AI companies that have released open weight models (Microsoft, Google, IBM, Mistral, DeepSeek, Qwen, etc.).
Training is limited on memory first (storing training data and weights) and computation second. Realistically you need to own or rent 2-8 H100/B100 devices or Google's TPUs.
The majority of workflows for a company providing AI capabilities are likely best solved by tailoring a system prompt for the chosen model, evaluating the prompt and model with tools like promptfoo, and then running it on a compute cloud provider (including AWS Bedrock). If the company is big/financially well off enough they could look at buying the hardware needed to run it on their own servers.
For other uses like agentic software development you'd need to spin up a suitable model on a compute cloud provider (or local hardware if the model is small enough) and then tell your IDE/editor to use that model. You would need some way of benchmarking and evaluating the models to see if they are capable of doing the tasks you need. -- There have been some tests done by people on YouTube that suggests that Qwen 3.8 27B is a decent model, but your needs may vary.
Most companies that build physical goods don't care for one second about their IT department other than how much money they can save per month, starting by outsourcing whole of it, thus they have little use for internal LLMs.
And it's across the industry, thinking banks, private banks, insurance, pharamcy etc don't outsource their IT, including development...
I believe US outsource even more than Europe on this matter.
That infographic says it uses 16-20% more fuel per unit of thrust. That's a significant amount more weight for an aircraft to carry, but because carrying extra fuel also burns fuel, extra fuel use would compound exponentially. This would probably mean significantly more than 20% extra fuel would be required for long-haul flights.
There's one plantation of under 1000 hectares in Papua New Guinea that makes biodiesel from coconut oil to power their own operation and vehicles on the island.[1]
But they don't export. On an island with few energy sources, the process makes sense.
This is more like an industrial process that self-powers from its own waste. Sewerage treatment plants do that - the process produces methane, often enough to power the plant. Not enough to put them into the gas business, though.
Amusingly, corn to ethanol plants in the US do not generally self-power. They consume natural gas to produce ethanol.[2] Net energy gain is maybe 1.3. Only subsidies keep this going. About 40% of US corn production goes into this boondoggle.
Biofuel as a side venture that makes a bit of profit / saves on a bit of fuel outlay has always been a thing in Australia also.
Currently, off the top of recent memory:
* Australia is exploring using canola to produce sustainable aviation and renewable diesel fuels, potentially creating 7,000+ jobs by 2030. - Landline (June 2025) https://www.youtube.com/watch?v=fIwsfxK6jw0
* Adelaide will turn 600,000 tons of forest waste into 140 megaliters of jet fuel per year, enough for 4.5M passenger seats. - (July 2026) - https://www.youtube.com/watch?v=z0u91oQ-hjY
The relative perspective on 4.5M passenger seats is that at one point in the recent past (possibly pre COVID) there were ~ 1M passengers in the air globally at any random time.
On the business investment and development side there are several operators, eg: JetZero Australia - https://jetzero.com.au/projects
I cringe thinking about more biofuel ideas getting forced through. Sure, let’s all spend a ton of petroleum energy on fertilizing, farm equipment, and processing, and raise food prices by using all the land for producing ethanol-corn, and jet fuel coconuts, just to get a slightly worse petroleum replacement fuel.
There's always the air-gasoline route. That's right! With enough chemical engineering and materials science, we can extract jet fuel from air and water. You extract out CO2 from the air, and H2 from water. You turn the CO2 and hydrogen into carbon monoxide, and then Fischer-Tropsch chemistry and some other magic gets you hydrocarbons out the other end! The problem is the economics of that process can't compete with oil from the ground.
> The problem is the economics of that process can't compete with oil from the ground.
Even factoring in emissions?
Given oil from the ground is adding long buried CO2 to and steadily increasing atmospheric insulation whereas fuel-from-air is using what is already there.
Is this another example of discussions about oil not factoring in the negative externalities to make it look good?
I think this is true to an extent, but it really depends on how much information the LLM has on a topic.
They’re amazing at maths because maths has been open source since day 0. They’ve great at algorithms for similar reasons. C bindings are a doddle because there’s so much prior art available. They’re terrible at using cutting edge features in languages because there’s not much data yet.
I think it’s fairly easy to predict what they’re good at on this basis. I’m not sure why they suck at UI design though.
I think they do OK, if you are fine with middle-of-the-road, default stuff. Also, I have found that it's important to provide guidance, constraints, and context to the LLM.
I have been working on a native iOS Swift app, for the last few months, and our team consulted an LLM (might be ChatGPT, might be Claude -I wasn't the one that consulted it), for graphic and interaction design.
Some context: We have written a 2.0/full rewrite of an existing app, that has been shipping for a couple of years. The 1.X version was designed by a professional graphic designer, but we didn't have him available for the rewrite, so we had to make do with our own guidance (questionable), and LLMs (also questionable, but for different reasons).
Anyway, As I started on the project, I kept getting designs that were, quite frankly, awful. They looked like generic Apple SwiftUI app screens, or worse, old Facebook app screens, with absolutely no relation to the current app. I had to basically ignore most of the graphic layout, and use only the interaction design.
After hearing complaints about "not making it look like the screens I sent you," I explained that they needed to train their LLM on the current design. There was no way that I was going to implement those ghastly graphics (It is so good to be able to do that. I'm retired, and working for free. In the old days, I would have had to take out a spoon, tuck in a bib, and eat shit).
They did that, and suddenly, their submissions were great, and I could start using them.
Thanks that’s an interesting anecdote to me because I’m also developing SwiftUI apps. Claude will create something that technically works but looks terrible.
I don’t have the luxury of a previously well designed app to train Claude on, but I’m finding just building the UI the old fashioned way then asking Claude to fix the bugs works to an extent.
I’m hoping that creating a comprehensive design system with good semantic naming will help.
That's a lot more than just visual design, as it incorporates things like branding and interaction.
You can keep iterating it, until you have something that works. It can be a fair bit of work, but worth it. In my case, I uploaded final PNG screenshots, along with detailed README files, explaining each screen. I also uploaded MOV files.
I use ChatGPT, which has a "memory," so I just declare the current state as "baseline," and it will always riff off of the design language. Not sure if Claude can do that.
Also, with SwiftUI, you have hard limits on what you can do, visually. I tend to use UIKit for my release products, because of that. I find UIKit to be a pain, but with it, and Autolayout, I can do almost anything.
Surely non-sandboxed build scripts are just a terrible idea.
Both Cargo and npm should look at what Swift Package Manager (SPM) is doing.
It’s not perfect but there’s a noticeable absence of supply chain attacks involving SPM, probably partly because it doesn’t use a mutable registry, but I suspect attacks are just more difficult. On the rare occasion a build script is involved it’s run in a sandboxed plugin.
If I was a pretty girl that gets stared at I think I'd want to record the ones doing the staring
People generally commit less crime if they think they're on camera
If they do commit a crime, at least there would be video evidence
Or if they just want to be creepy, I'd send those stares to their wife
Considering motivated perverts would use hidden cameras, as opposed to cameras literally on their face, I think the asymmetry here is in favor of recording the perverts rather than vice versa
I think there's a difference between brandishing a weapon and brandishing a recording device
Unlike the sword, the camera itself does not incite deeds of violence
My understanding at least is that cameras (or even signs) reducing/preventing crime by _some_ amount has been independently corroborated across multiple studies
Maybe I lost too much faith in society. But in a world where everyone is recording each other, I think people who are able to act with impunity wins, not necessarily those with higher moral standards.
reply