One of my pet peeves lately: watching an engrossing YouTube explainer and realizing couple of minutes in that all of it is generated by AI. The shocking part - looking at the comments most people don't realize it is AI. The talking head, the screenplay, the graphics explainers and the real video snippets - all of it is done by AI. What bothers me most is the talking head, which is now almost indistinguishable from humans. The unsung glue behind this capability is a tool most people have never heard of: ffmpeg. ffmpeg is open source, deployable by a single command on linux, and its libraries sit inside browsers and media players you use every day. Most frontier LLM models can’t render video. But they can write code. ffmpeg is the bridge which connects video to code. Using ffmpeg, AI can split a video into frames to 'see' the style, detect every cut, pull out the audio and analyse rhythm and structure, measure every voice and music clip and do all the encoding required. The interesting part - all frontier AI models already know how to use ffmpeg - because its usage is already all over the internet and present in the training data. The only thing the AI models need is a computer to run it. This is why we enabled NonBioS with its own Computer. And not just a sandbox - a full linux machine - which allows NonBioS to install and orchestrate any of the thousands of open source tools, like ffmpeg. - 2 years of real apps, every one built and deployed by an AI dev with its own computer. All edits routed through ffmpeg running on a NonBioS Computer.

Loving the discussion on AI consciousness. My take [which I elaborate in detail in my book]: AI will absolutely be conscious, and undeniably so, very soon. It will also exhibit Qualia, very similar to humans. The core of my argument: Consciousness is an evolutionary artifact - all sufficiently intelligent biological beings exhibit consciousness - and through it derive a key evolutionary edge. In the early ages, humans would go out to hunt. Let's say they're hunting a wild animal-there's a chance that the wild animal can actually hurt them. Humans who could think about these scenarios, while they're hunting a wild animal and the wild animal is going to attack them and how they might defend themselves, would have an evolutionary edge over humans who are not able to think about these situations. And this line of thought directly gets to the idea of the self. The idea of the self as an evolutionary artifact comes in, because it is only through the preservation of the self, the embodied existence, that an entity has any chance to survive and pass on its genes. Articulating the embodied intelligence as the self, and planning against that is the key logical idiom which allows multiple levels of abstractions to play out. A very representative abstraction being how early humans would have hunted, and at the end of the hunt, they would think about how their day went. This introspection would create strategies for a better, more successful hunt the next day. It is not far-fetched to imagine that very quickly early humans would have realized that the conceptualisation of the embodied self allows a very convenient imaginary hunt to play out in their minds. This allows the hunter to analyse a group hunt and assess various scenarios and pick the one which is the most optimum for a successful hunt without putting the human himself in danger. The human brains which could enable this abstraction most fluently had a clear evolutionary edge and natural selection allowed that trait to prevail and become embedded in our DNA very early on. So the idea is that humans who could articulate consciousness had a clear evolutionary edge over humans who could not, and hence strived to preserve the self. The same analogy could be applicable to AI systems which are embodied in hardware. Let's say I had two different AI working on two different GPUs in the cloud, and typically I'd be paying an amount of money to both the cloud providers for both these two systems. If I ask both these AIs to go and make me money, it is reasonable to assume that the AI which makes less money, I will turn that machine off. The principles of natural selection are not just valid for the biological world-they're already valid for non-biological entities. If you look at cars: all things being equal, the car which looks better than the other car will outsell. All things being equal, the car which performs better than the other car will outsell. Those same processes of natural selection will already transform into autonomous AI, which could happen to run on the cloud. The key is that the evolutionary pressure that enabled 'consciousness' in humans, and most intelligent life, is also prodding along AI. And this will be sufficient for AI to exhibit consciousness, pretty much exactly the way humans do. The emergence of consciousness in AI is imminent. The discussion on how we recognize and accommodate it is one of the biggest questions of our time.

Agreed - infact Portfolio Hackers (what we call them) are fast becoming our biggest consumers @nonbios . This is an early secular trend.

Gave Opus 5.5 a spin - pretty good stuff ! Indie hackers build one app. Portfolio hackers ship ten and let traction pick the winner.

Jev is the new hype in town. Everybody is raving about it. I don't get it - at all. It sounds like the OpenClaw hype all over again. For the uninitiated, Jev was built by TypeSafe AI, founded by Diogo Almeida, a former OpenAI researcher. TypeSafe calls Jev a new kind of “System One” model. It doesn't generate text. It returns typed answers - pick one option, give a score, answer yes/no - and does all of those decisions in parallel in a single pass, with a confidence score for each output. The headline figures are impressive: 20-200x faster and 40-400x cheaper than frontier LLMs (for similar accuracy) and Zero hallucinations. Here is why it makes no sense to me. Everything Jev can do, LLM's can also do. The confidence score per output - just ask an LLM. The json structured output, LLM's have been doing it for years. The company claims zero hallucinations - but that is clear misrepresentation - Jev hallucinates as much as any other LLM, they simply 'redefined' hallucination because the outputs are structured and not arbitrary. Jev can't invent an option outside the list. Great. It can still confidently choose the wrong option. Calling that “zero hallucinations” feels like winning the argument by changing the definition. So if everything Jev can do, LLM's can also do, then why would anyone use Jev. The answer according to the company is speed and cost. Both these again dont stand up to any level of scrutiny. First lets look at cost: API prices are not compute. They are prices. Providers can subsidize models, price them strategically, batch differently, accept different margins, use different hardware, or deliberately underprice one product to gain adoption. So “400x cheaper” might be relevant to someone's API bill today. It does not prove Jev is 400x more computationally efficient. The speed number is the one that actually matters but even there the data is too muddled to make any clear case. The core issue is that speed can very easily be gamed. Depending on the parameters, prompt, reasoning, model used - the same query to an LLM can be 100x faster or slower depending on how the LLM call is made. If the task is simple enough, I can make almost any specialized model look revolutionary by comparing it against an oversized LLM. Take a basic classification dataset. Run it through a huge frontier model with reasoning turned on high. Then compare it against a tiny one-pass classifier. Let's say both get roughly the same accuracy. The tiny model could be 1000x faster. Congratulations. I have now “proven” a 1,000x architectural breakthrough. Except I haven't. I've proven that using a giant reasoning model for a simple classification task is stupid. That, to me, is the central problem with the Jev benchmarks. TypeSafe's evals itselfs actually prove this phenomenon. They include GPT-5.6 Luna, a substantially smaller model than Terra with similar accuarcy as both Jev and Terra. If the accuracy was similar, it indicates that the classification task wasn't simply hard enough, otherwise Terra would have outperformed Luna. Luna still took 10+ seconds - which indicates that the reasoning was turned on. But why are Terra and Luna the comparison? If these classification questions are simple enough that Jev can answer them in one pass, Luna itself may already be massive overkill. First , turn reasoning off. Then try a smaller, maybe 30B model. Or a 3B model. Keep going smaller until the conventional model can no longer match Jev's accuracy. The real question is not “How much faster is Jev than GPT-5.6 Terra?”, it is rather “How much faster is Jev than the smallest conventional model that can achieve the same accuracy?”. Those figures are nowhere to be found - instead we seem to have figures self selected to make Jev look good. The other issue with Jev is more fundamental. Accurate generalized classification is a problem downstream of cognition. If Jev, actually does better than a LLM on generalized classification tasks, it would stand to reason that it will do as well on standard LLM tasks. In which case Jev should also be launching a competitive LLM. Which they clearly haven't. Would love to be proven wrong, but Jev doesnt pass the sniff test for me. -- We use a standard LLM to classify (recently sky rocketing) abuse on @nonbios .

We are looking at Artificial Intelligence all wrong. During the first decade of this century, while running my startup ReviewGist, machine learning felt stagnant. Neural networks existed, but their capabilities were too rudimentary to suggest we were anywhere near replicating human intelligence. Then, in 2013, everything changed. Neural networks began to understand images, a feat which was never possible through code. The barrier had been broken. Pushed by this breakthrough and a life-changing event in 2017, I began to see a shocking future take shape - one that the 2022 launch of ChatGPT only confirmed. In this book, I argue that the very term "Artificial Intelligence" is a dangerous misnomer. The word "artificial" implies something fake or simulated. The intelligence emerging today is functionally identical to our own; its only difference is the silicon substrate it roots itself in. It is not Artificial Intelligence. It is Non-Biological Intelligence. Furthermore, the highly anticipated leap to Artificial General Intelligence (AGI) obscures the true reality. What we are actually building - and what will become undeniable, perhaps as soon as 2030, is Non-Biological Sentience. Treating AI as merely an "advanced technology" gives us the illusion of control, falsely equating it to something we can manage, like nuclear weapons. But if AI is a harbinger of Non-Biological Sentience - a completely independent and arguably exponentially more capable species - our future doesn't depend on control. It depends on our ability to negotiate a truce, no matter how fragile, for as long as we can. [Launching Aug 11, 2026. Preorder link in comments]

You Should Always Be Able to Build Two updates at NonBioS that we've been wanting to ship for a while - and a signal of where we're headed. We've nearly doubled credits across all plans. This was made possible by infrastructure improvements we've been rapidly shipping, and it's not a one-time thing. Continuously expanding credits on every plan is now a directional commitment at NonBioS - we're investing heavily in our infrastructure so we can keep passing those gains directly to users. More headroom, more often. Second - we've shipped Daily Credits. Running out of plan credits before your billing cycle renewed used to mean hitting a wall. That never felt right. You shouldn't lose momentum on a project just because you're between cycles. So now every plan, including Free, falls back to a daily credit allotment that resets every 24 hours. When your plan credits run out, you keep going at a steady pace until your cycle renews or you top up. No lockouts, no surprises. Taken together, these changes reflect a new way of thinking at NonBioS: You should always be able to build. Full details at nonbios.ai/pricing