And I think it’s important to link to them here. These aren’t casual skeptics – they’ve thought about it and concluded AI is fundamentally harmful.

• Every Reason Why I Hate AI and You Should Too by Marcus Hutchins (the security researcher who stopped WannaCry)
• I Am An AI Hater by Anthony Moser

I mention further down that there will be “AI vegans.” These people are already here and their concerns are worth reading even if you don’t fully agree.

No. And this surprised me.

Firefly is trained only on licensed content by Adobe. This seems like a great decision for them, especially from a quality control perspective.

Our approach to generative AI with Adobe Firefly

We only train Adobe Firefly on content where we have permission to do so.

Phi-4-mini claims to be trained on high quality data like licensed books and academic sources. This is a very good, tool call capable model that can run on a laptop or phone, so this opens up a ton of local AI use cases where you use your own energy to do AI tasks with a more ethically trained model.

phi-4-mini on Ollama

Phi-4-mini is a lightweight model built upon a blend of synthetic datasets, data from filtered public domain websites, and acquired academic books and Q&A datasets. The goal of this approach was to ensure that small capable models were trained with data focused on high quality and advanced reasoning.

So it seems very possible to build SMLs and maybe even LLMs in an ethical way with enough effort and will.

However, all the “big models”, frontier LLMs, are in a grey area today. Courts are now ruling, but in contradictory ways: the UK High Court ruled (Getty v. Stability AI, Nov 2025) that AI training does NOT constitute copyright infringement because trained models don’t “store” copyrighted works. Meanwhile, a Munich court ruled the opposite (GEMA v. OpenAI, Nov 2025), finding OpenAI violated German copyright law by training on song lyrics. In the US, the NYT v. OpenAI case is progressing but hasn’t gone to trial yet. And the music industry is settling into licensing deals rather than waiting for courts.

Extra-legally, it can feel like theft for sure. The vibes are mixed and that is worth acknowledging. I don’t believe the law will actually help us here. The egg is already scrambled and courts across jurisdictions can’t even agree on the basics. Extra-legally, pressure can always be applied towards companies that they behave in a more ethical manner. Sure. We should always strive for that. We don’t have to accept raw capitalism. We are always making trade offs, and that will continue.

While I have read more than one person online trying to say it is, I have found no reason to believe that it is.

Matrix transforms + a giant embedding space seems to be the main magic here, and that’s math.

It helped me to learn more about what is actually going on. Checkout these links:

• Transformer on Wikipedia
• Transformers, the tech behind LLMs | Deep Learning Chapter 5
• 👉 The moment we stopped understanding AI: AlexNet

If you watch only one explainer video, the AlexNet one is the best to really explain what is going on inside these things and how we got to where we are today.

The internet does not have one economic model. Advertising is Google’s economic model (and then Facebook copied it as well). Saying “advertising is the economic model of the internet” benefits Google and Facebook, it’s the story they want told. There are other economic models working. Those will remain. Advertising may take a pretty big hit, for sure.

Definitely not.

And this is one reason I am working hard to enable local AI workflows on phones and laptops. I want to make it easy to bend this tech to benefit us, not bend ourselves to benefit it. It’s more important than ever that we are in control of our data, of our context.

First, I hate that you made me type “web3” on this here website.

Second, web3 is a lot different. My coins get more valuable if you buy a coin. It’s that simple. So I need as many people as possible to buy coins, so I can buy low and sell high. Stable coins might have utility, a lot of adjacent research and math is useful, IPFS is cool, sure, but overall it really seems mostly like a way to make new speculative assets.

AI is not like this at all. My AI tasks don’t start working better, or become more valuable if you use AI too. Also, most of the big AI companies are still losing money because of how expensive all of this is for them, though there are exceptions – Midjourney is profitable with zero outside funding, and Microsoft is making real money on Azure AI. Most AI companies are “valuable” today because of future profits, not their profits today. They have cash flowing through them, but most are not capturing much of that cash… instead they are spending money to watch all their revenue flow out. The chip maker is doing quite well though, if you haven’t seen.

So it’s just not the same. The parts that feel the same are probably just general hype cycle dynamics.

Yes, probably.

I’ve read quite a bit about this now and the honest answer is: the reporting doesn’t do a great job of putting things into perspective. A single ChatGPT query uses roughly 5x more electricity than a web search, but what does that actually mean compared to, say, the energy I use making coffee or driving to the store? The articles I’ve found are better at describing the scale of data center growth than at helping a normal person understand where AI fits into their own energy footprint.

That said, the trajectory is concerning. MIT Technology Review reports that by 2028 more than half of data center electricity will go to AI, and no company reports AI-specific environmental metrics. Brookings reports that a typical data center uses 300,000 gallons of water per day for cooling, and large ones use 5 million gallons – and water cooling demand may increase 870%.

The real concern isn’t today’s usage but the growth trajectory. AI usage is expected to double or triple in the next few years, and infrastructure decisions being made now will lock in environmental impacts for decades. Efficiency improvements are real (newer chips, better cooling) but are being outpaced by demand growth.

Listen, I am still bitter about “cloud computing.”

You are correct, it is not “intelligent.” However, you can’t always win the messaging wars. I’ve moved on.

Related: https://solarpunk.moe/@alilly/114928042375589900

An agent is an LLM or SLM with possible tool/function calls, running in a loop. The LM can generate a spec to call a tool, another program calls that tool, then the previous conversation + the return value of the tool is fed back into the LLM. Repeat. Sometimes there is a function call to end the loop, the LM can generate the spec to call that to finish the task.

”Tools in a loop.”

There are many other definitions of “agent,” but this is the one I like best right now.

RE this article by Simon Willison: Tools in a Loop.

Related: https://mastodon.cloud/@jasongorman/114595098303670564

It is more nuanced. Next word prediction is very important for these products, but there are a few more things going on.

One thing that is worth watching is this video about Google’s Alpha Geometry project and how much “not AI” there is in that system. You don’t need to understand all of the geometry to understand that the AI part of the program isn’t even half of the whole deal.

Another example is QueryGPT. A specially trained LLM can generate SQL from a plain English query. Then a normal database system will run the SQL and return the results. And, if one wants, the return value from the database could be fed back onto an LLM to generate a more “human friendly” response.

Generative AI (different types of fancy prediction) output is very useful as an input into another system. ChatGPT debuted as just Generative AI without much else, it would spew back text to you and that was it. But today, all of the major AI products are a series of workflows and pipes, where one or more of the steps is generative.

So, yes and no.

And hopefully you can start to imagine how “generating statistically likely text / code to feed into something else” could be useful sometimes.

Yes. Google search pretty much sucks right now.

AI research tools can be much better at surfacing the long tail. LLMs themselves have nothing to do with search, but “AI tools” and “agents” which might use LLMs to generate search queries, filters, etc can do a better job at searching than the average person. It feels to me like we are just beginning to see how LLMs and SLMs can help us improve our searching.

Related:

• I am disappointed in the AI discourse by Steve Klabnik
• Local Open Source GPT Researcher

While working at Microsoft I heard a lot of Bill Gates stories. 😅 And while they might just be legends, one of them I remember and is related.

It was said that when Bill needed to learn about some difficult topic, he would pay a team to setup and video record lectures on the topic at top Universities. That team would then synthesize those recordings into a compressed curriculum for him, deliverable in a single binder. Then he could review that and quickly become a pseudo expert. And what a smart idea!

AI research tools can assemble a single folder of compressed information for you today and this works well. And this is a new super power. You can do what Bill Gates did (or maybe didn’t do, but was said to have done).

Giving the LLM some research input and having it generate a distillation or summary is where things might go wrong. It could generate nonsense, sure. Having the folder of resources is the most important part of the final artifact, not the generated “human friendly” summary.

Yes.

In my experiments, it does as well as an average person. And I say this confidently. I’ve worked with enough programmers that I think you could easily hit the average with today’s tools.

Now, to be clear, the average is a pretty low bar, so this isn’t as exciting or damning as it might sound. Today’s AI coding tools seem exactly like an unreliable coding intern who is in a hurry to go home. Which I guess is an achievement for humankind. That said, every 3-6 months the quality and reliability of these tools really is increasing, and I do expect them to keep getting better.

Code feels easier to accurately generate than normal language to me, because of its limited grammar. And it can be tested to prove that it works. So I think it’s about the loops of tools that go from generate to test to remove, etc. Justin Searls has written that TDD is more important than ever precisely because AI coding agents need verification to work well. He also built a project in three hours that would have taken weeks, making it worth doing at all.

There is real research now. A large RCT at Microsoft, Accenture, and a Fortune 100 company (4,867 developers) found Copilot increased completed tasks by 26% – but almost all the gains went to junior devs (27-39%), while senior devs saw only 8-13%. Meanwhile, a METR study of 16 experienced open-source developers found that AI users were actually 19% slower on their own repos – and the developers believed they were 20% faster. That perception-reality gap is worth sitting with.

Related: https://wandering.shop/@aesthr/114592630789058368

Also: Sprinkling Self-Doubt on ChatGPT by Justin Searls.

Yes.

Humans have done that for a while. Now robots will do it in a more scalable way.

Related: https://xoxo.zone/@microwavenby/114672517338884522

Also: https://neilmadden.blog/2025/06/06/a-look-at-cloudflares-ai-coded-oauth-library/

Also: https://forum.cursor.com/t/cursor-yolo-deleted-everything-in-my-computer/103131

The most successful projects I’ve worked on have been where I was fixing some awful, existing code. So if you enjoy fixing broken projects, this is your heyday.

This needs to be split into two answers now.

Reproducing something visually, pixel for pixel: mostly no. Posters and graphic design are not code or bitmap images. You can generate bitmap images using AI tools, but that is not quite the same as “design.” Simon Willison always has each new model generate a pelican riding a bicycle and the results are informative.

Creating designs and working apps just by prompting: very much yes. Figma Make can turn text prompts into working app prototypes and designs, inheriting your existing design system, with no coding required. If you want a website or app prototype that is as good as the average, then yeah, you can get that from a prompt today.

And you can now generate 3D models from text prompts. Tools like Meshy, Tripo AI, and Luma Genie can generate textured, export-ready 3D meshes in 30-60 seconds from a description. These cover modeling, texturing, retopology, and rigging. (Comparison of text-to-3D tools for 2026)

This area is moving very fast.

Some are, with real numbers now.

Midjourney generated ~500Minrevenuein2025withabout100employeesand * *zerooutsidefunding * * − −genuinelyprofitable.[OpenAI](https : //finance.yahoo.com/news/openai − cfo − says − annualized − revenue − 173519097.html)hit 20B annualized revenue but is still burning ~$8B/year. Anthropic topped $9B run rate. Cursor crossed $1B ARR but may be spending nearly 100% of revenue on API costs. Microsoft Azure is growing 39% year-over-year, driven largely by AI.

So there is enormous revenue, but actual profitability among pure-play AI companies is rare. Midjourney is the standout.

That said, a MIT NANDA initiative study found that 95% of generative AI deployments fail to generate measurable business value. The money is flowing but the value is concentrated.

Duolingo’s earnings are a window into the disconnect between the vocal minority who complain about AI online and the value businesses & people are getting out of it…

https://mas.to/@carnage4life/114993379869191876

Also: Some travel advisors are using AI to help plan trips and boost business

Maybe not.

And yeah, that is dystopian. I am not excited about this, but if current trends hold I don’t see how you can be 100% sure.

• Deepfakes, Scams, and the Age of Paranoia

I don’t think this is actually the right question. Many programmers move into roles where they write less code and spend almost all their time reviewing code. This is natural over time. And this could happen with AI: code review instead of code gen for code experts. Copy editors are the same: they are good at editing, someone else generates the copy.

Also, photograph didn’t kill 100% of painting, but it definitely made painting an extra special, rare thing. You can always keep painting, but it might not be the dominant job anymore.

Seems likely.

It seems to me, if you are an expert, then you are less likely to atrophy. If you are not an expert yet, and you don’t put in the effort, then you are not building any “muscle.” This is true today: if another human does the hard work for you, then you didn’t learn anything. And so that seems likely to be true with AI.

There is now real research on this. An Anthropic study (Jan 2026) had 52 junior engineers complete coding tasks with and without AI. The AI group scored 50% on a comprehension quiz vs 67% for the hand-coding group – about two letter grades lower. Debugging skills were hit hardest. And the time savings from using AI were not even statistically significant. As they put it: “AI-enhanced productivity is not a shortcut to competence.”

Meanwhile, a Stanford study found that employment for software devs aged 22-25 has declined nearly 20% from its peak in late 2022. If juniors are being hired less, the pipeline for developing senior talent narrows.

Related: In the Long Run, LLMs Make Us Dumber argues that over-reliance on LLMs weakens cognitive abilities by removing beneficial mental friction.

If you are going to use AI, your best bet is probably to ask the AI to help you become an expert, and not to just give you the answers.

Not every product.

Checkout https://procreate.com/ai, for example. My prediction is there will be a few apps that either intentionally stay out of AI, or AI just never is a good fit for.

Meanwhile, some people are frustrated that AI is being added to products they don’t want it in. One person is considering blocking all browsers except Chrome because it’s the only major browser without built-in AI summarization.

No, not everyone.

There definitely will be “AI vegans” and with diverse views about why they are avoiding AI, just like there are diverse views about avoiding meat.

No.

There are very capable open source models, so you’d have to delete code permanently from the universe, and we know that is impossible.

I mean, sure, every government could outlaw AI and the open source models go underground. But this feels like a fantasy to me, and not really worth considering further.

Yes, it will not feel great for some people.

I am personally not excited to become a robot engineering manager… and it definitely feels to me that that will be a job. I think the main unknown is to what degree things will change. Will there be a few people that change over to manage robots, or will the majority of knowledge workers change over? If I had to guess today, I’d guess majority.

Related:

The thing that keeps coming up as I talk to people about AI in their workplaces is how dehumanizing it is. It’s dehumanizing to ask a machine to do something, and then have to correct it over and over; it’s dehumanizing to be told to read something that involved little to no human effort to make.

https://mstdn.social/@aworkinglibrary/114659560902662745

Justin Searls wrote about The Generative Creativity Spectrum: when creativity’s value is internal (writing essays for yourself), AI diminishes it. When it’s external (producing output for others), AI can help. Programming sits uncomfortably in the middle – AI dramatically increases productivity but diminishes the engaging problem-solving that made coding rewarding.

And steipete wrote about the flip side: AI can be addictive. “Just one more prompt” mirrors unsustainable work patterns. AI restored his passion for building but created a compulsive loop.

AI agents can generate code and features so quickly that I may not have time to actually try them out. I have a side project right now where there is something waiting for me to test, to actually sit down and use it, and I haven’t had time. The code was generated faster than I can evaluate it. I am the bottleneck.

This is a weird inversion. The hard part used to be writing the code. Now the hard part might be having enough hours in the day to verify, test, and actually experience what was built.