"Software will be a commodity, like electricity, and it will run directly in the LLM layer." What does that even mean? I’ll break it down in detail (long post incoming 🍿) and once you finish there won’t be any middle ground. You’ll either reject the idea completely (and be wrong) or you’ll walk out a convert and I’ll welcome you to my cult 😂 Yesterday’s post, where I said I can see this coming very clearly in the near future, stirred up a lot of buzz and questions. So I’m going to flesh out the idea I’ve been sharing here (and with anyone who doesn’t get bored of me outside X, I’m tired of myself at this point) for over a year. Let’s get to it. 1. ‘All software will run in the LLM layer’ First, the obvious. Even today there are people who, unbelievably, still deny we’re heading toward a world where all software (frontend, backend, infra, etc.) is created with plain natural language. In reality, we’re basically there already with LLMs like Claude, Grok, or ChatGPT and tools like Cursor, Windsurf, or Lovable. Any decent dev I know doesn’t write a single line of code anymore. They speak in natural language and direct swarms of agents to build the whole app, then they just review it (and argue and beg 🤣) so the AI delivers their vision. That part is trivial. If you’re not there yet, it’s because you don’t want to be. The next leap is way bigger, and it’s where even people deep in AI look at me with skepticism. We won’t need that layer either. No Python, no HTML, no C++, no CSS, nothing. Everything (the whole software layer, the logic, the persistence or memory (call it a database if you want, but it’ll be much more than that), will ‘run’ directly inside the LLM. Elon Musk summed it up in a line that’ll go down in history: "Any input bitstream to any output bitstream." If you know tech, you probably don’t need more. You already get it. Just in case, here’s a quick example. Today, if we want to build an upscaler like Magnific, we need: - Backend logic, the code running on the server (in our case, Python) - A database (Firebase, Postgres, Supabase, MySQL, etc.) - The frontend layer, what the user sees: HTML, CSS, JavaScript - The infra layer: cloud GPUs where the magic runs and where we call workflows via APIs What a circus! And sure, you can pretty much build this with Cursor + Claude already. But programming languages were designed for humans. Future AIs won’t need them. They’re redundant. A parameter we should remove. An inefficiency. Do you really think all of this is necessary? Do you really think we’ll still be doing things like this in 20 years? No. A future LLM, one that’s hard to imagine today but that we’ll see sooner than you think, could take a screenshot of today’s Magnific and, with just that image and a mouse click or drag, infer what it needs to do next. INFER WHAT IT NEEDS TO DO NEXT. And I don’t mean it’ll secretly build a frontend, backend, and database under the hood (that’s medieval). I mean it will literally generate the next visual frame to show the user. That could be the same screenshot with the slider moved 1 pixel to the right and the displayed image with a slightly more detailed upscale. Just like that. ‘But… but… but… Javi… where does the logic happen? Where’s the program doing that?’ Directly inside the LLM! The same way today’s models can take an image and generate a full video with decent physics, they’ll understand what we currently call the ‘business logic’ of any application, no matter how complex. "Any input bitstream to any output bitstream." Logic emerging from the model itself, not from hand-crafted code. Now you’ve probably got it. If not, take a minute to chew on the idea before moving on to the next point, don’t let it get stuck in your throat. 2. ‘Software will be a commodity, like electricity’ And not just software. All digital entertainment too (movies, videogames, etc.). In real time. Most digital tech value will concentrate in very few companies. The ones that win the multimodal LLM race and the ones that provide the infrastructure they run on. Picture a world where saying ‘I want a SaaS that does this’ or ‘make me a movie in this style with my dog as the lead’ gets you an instant result from an LLM at a quality far above today’s best productions. I basically think all logic and visual layers will run on ultra-advanced LLMs we can barely imagine today. It won’t make sense to build apps, sites, or entertainment the way we do now. The ability to do it will concentrate in companies with the best LLMs and the compute to run them at scale. We’ll pick AI providers mostly on price and much less on features or capacity. Think electricity companies today, or PS vs Xbox if they lock in some interesting IPs. 3. ‘But Javi, they aren’t deterministic and they don’t have memory. They’ll never build decent apps like this’ Calling it impossible is short-sighted. Of course they’ll have persistence and memory. Just not in the simplistic ‘like a database’ way we use today. 4. ‘Ok, but how would you even build an LLM like that?’ Carefully 😂 Mainly with synthetic data. Here’s an obvious idea that’s still pretty revolutionary. Since LLMs and agents are close to generating end-to-end apps (backend and frontend), we can instrument them to build in a closed loop, with explicit objectives and reward functions, logging UI states and inputs. Those trajectories can feed training for future foundation models. If I can think of this with my limited knowledge (even if no big company has said publicly they’re doing it), I can assume much brighter minds are already on it. 5. ‘If this turns out to be true, how do you get an edge over the non-believers?’ You’re already ahead if you’re right about something like this. For now, my move is positioning through investments, figuring out which companies will concentrate most of the power by locking down GPUs and the tech. I’ll drop a link later in the thread, though it may be short-sighted and focused on what exists today. Best you can do is follow the game day by day and stay sharp. How’s that landing? Any questions, happy to answer in the comments.

bro...

Airbnbs, and I mean truly premium Airbnbs, used to be places where you’d pay less than in a hotel and get a lot more, like being able to check in early and check out late, etc. Now they are far more expensive that 5 star hotels and make you take out the rubbish under threat of a fine, won’t let you in until 4 pm, kick you out at 10 am, and give you a long list of instructions on how to leave the place, even though you’re paying €300 for so-called cleaning and €700 a day for the place itself.

It goes up up up until it goes down down down.

🔥 I see it clearly that the future will be exactly like this: All software will be absolutely liquid and run directly in LLMs. Software will be a commodity, like electricity. Seeing this in advance can give you a huge opportunity 🤑 I'll be there to take it.

LOL. Is not that easy to kill legacy models. We will offer "Magnific 1.0" like forever, just in case. We won't try to make you swallow anything new without you really choosing it.

Going some days to Ericeria! 🇵🇹 Who's there? Do you want to grab a coffee with me? Thank you very much!

Tomorrow news: Lovable killed OpenAI. And Video Killed The Radio Star!

⚡ GPT-5 just killed Loveable. Eventually, all layers of software will run directly inside LLMs. The massive amount of power that AI companies with GPUs and the ability to keep releasing new LLMs will have will lead to absolute monopolies.

⚡ Can you believe this is being generated in REAL TIME? Videogames as we know them are going to change forever. There will be just ONE video game. One that can create ANYTHING you can imagine.