Skyfall AI's Next Steps with Sam Pasupalak

Most AI startups are building harnesses on top of LLMs. The entire AI industry is celebrating how the ‘AI software engineer’ is the breakthrough of our generation. There used to be a time when Yahoo and AOL were the undisputed kings of the internet by being the best curators and organizers of existing information on the web. However, they lost relevance very quickly because they were optimizing for the easy win (curating information manually via humans instead of indexing and searching all the websites). Google’s page rank algorithm quickly became the dominant search tool and paved the path for Google to be the ultimate winner in the race. 30 years have now passed and we are at a similar junction of time again. Instead of focusing on the low hanging fruits of building ‘AI co-pilot’ experience, we want to build completely autonomous businesses end-to-end. We are building the foundation layer of super intelligent enterprises that require little to no human intervention, which we call ‘AI CEO’. The benefits of this are quite obvious and profound - increased productivity for the entire human race so that all of us can focus on our own passions in sports, arts, creativity or spending time with our loved ones. I shared more about Skyfall AI’s next steps in the The New York Times article that came out this weekend. Thank you Sarah Kessler and Andrew Ross Sorkin for the feature.

The honest truth about AI in your startup that nobody is saying out loud Every week I see another post claiming AI will 10x your startup velocity. Some of it is true. A lot of it isn't. Here's what I've actually found after building AI features into real products over the last year: AI is genuinely fast for structured, repetitive tasks. Auto-tagging, summarizing, suggesting edits, scoring inputs. If you can describe the task in one sentence, AI can probably do it reliably. AI is unreliable for anything open-ended. "Give the user the best answer" is not a well-defined task. When outputs are vague, users stop trusting the feature. And once trust breaks, they never come back to it. The real cost isn't the API. It's the review loop you didn't build. When an AI feature gets something wrong (and it will), you need a way for users to flag it and for you to see patterns in the failures. Most founders skip this. Then they spend weeks debugging by guessing. Speed is real, but scope creep is realer. AI tools let you prototype fast. That speed makes it tempting to keep adding. "Let's also make it do X." Every addition needs its own reliability threshold. I've watched founders build 6 AI features in a week and ship 0 of them because none were actually production-ready. The founders getting real traction from AI right now are picking 1 high-friction user moment and solving it well. Not building AI-powered everything. If you're adding AI to your product, start with the question: "What does the user hate doing that we could just do for them?" Everything else is noise. What's one thing about AI tools that matched your expectations, and one that didn't? Follow @satyyyaamm for more.

🧠 What comes after the transformer? Transformers have powered the LLM boom, but new approaches are emerging to make AI faster, more efficient, and potentially smarter. From smaller models and new attention mechanisms to generating text all at once and moving beyond language-based reasoning, startups are exploring what the next generation of AI could look like. If you're interested in knowing where LLMs may be heading next🚀read the full article linked below👇 https://lnkd.in/dZadnbfT

The AI race is no longer just about who builds the smartest model. It's becoming a battle over who gets to shape the future of AI itself. Over the past year, I've noticed something interesting. Some of the most capable AI models are no longer coming only from closed, billion-dollar AI labs. Open-weight models are improving rapidly.🤔 They're cheaper. More accessible. And they're giving developers and businesses far more control. That shift is making some of the biggest AI companies uncomfortable. Not because open models are necessarily better today. But because they change the economics of AI. If businesses can run powerful models on their own infrastructure at a fraction of the cost, the competitive advantage of proprietary AI starts to shrink. That's why this debate caught my attention. It isn't really about one Chinese model or one American company. It's about a much bigger question. Should innovation be protected by limiting competition, or accelerated by encouraging it? History offers an interesting lesson. The technologies that changed the world the most weren't always the ones kept behind closed doors.☝🏻 The internet grew because its core protocols were open. Linux became the backbone of countless systems because developers around the world could improve it. PyTorch became an industry standard because the community helped it evolve. Open ecosystems don't just create competition. They create participation. Of course, national security, privacy, and safety are real concerns, and they deserve serious attention. But those concerns should be addressed directly, not by assuming openness itself is the problem.🤷🏻 To me, the most exciting part of AI isn't that a handful of companies can build incredible models. It's that millions of developers, researchers, startups, and students can build incredible things on top of them. That's where innovation compounds. And that's why I think the future of AI won't be decided only by the companies with the biggest budgets. It will also be shaped by the communities with the biggest ideas.🙂↕️ #OpenSourceAI #ArtificialIntelligence #Innovation #AIDevelopment

A small AI data startup just posted eye-popping growth. Micro1 makes synthetic data and labeled datasets that other companies use to train AI models. In just eight months, its yearly revenue run rate jumped from $100 million to $500 million. The company keeps 60 to 70 percent of that as net revenue, with margins as high as 80 to 90 percent. Founder Ali Ansari says Micro1 will not sell its data to Chinese AI model makers, a stance that sets it apart from some competitors in the space. Micro1 raised its Series A last September at a $500 million valuation, and reports suggest it may have just closed a new funding round at an even higher price. This is a good reminder of how much money is flowing into the unglamorous but essential work behind AI: the data that trains the models.

The strongest AI coding model is not always the best model to use. This chart from Artificial Analysis shows why. The vertical axis measures coding performance. The horizontal axis measures API cost. And the interesting part is not simply which model reaches the highest score. It is *how much additional performance you get for every additional dollar you spend.* Some models reach very strong coding performance while staying relatively inexpensive. Then something changes. As you move toward the highest scores, every additional improvement can become dramatically more expensive. That creates an important question for anyone building with AI: Why use the most expensive model for every task? A production AI system might be better designed like this: 🔸 Cheap models for classification, extraction, formatting and routine coding tasks. 🔸 Mid-tier models for most development work. 🔸 Frontier models only when difficult reasoning, architecture or debugging actually requires them. This is why I think *model routing* will become increasingly important. The goal should not be: "Use the smartest model." The goal should be: "Use the cheapest model that can reliably complete this specific task." At scale, that difference can completely change the economics of an AI product. The next competitive advantage in AI may therefore come not only from having access to better models, but from knowing *when not to use them.* 🚀 Building an AI or technology startup? Join us in San Francisco on September 2 for D4SF Pitch Night Competition. Pitch what you're building, meet founders and builders, and see what other teams are working on: https://lnkd.in/gkWax_Up What matters more to you when choosing an AI model: maximum performance or price/performance? #AI #ArtificialIntelligence #OpenAI #Claude #Anthropic #Gemini #Coding #AICoding #LLM #AIModels #Developers #SoftwareEngineering #Startups #Founders #ModelRouting #AIEconomics #SiliconValley #SanFrancisco

AI just got dramatically cheaper, and I don't think enough people are paying attention. OpenAI recently slashed the price of its cheapest GPT-5.6 model tokens by 80%. Input tokens went from $1 to $0.20 per million. Output tokens dropped from $6 to $1.20. That sounds like a pricing update. It’s actually much bigger than that. A year ago, running an AI feature at scale meant constantly asking: “Can we afford the inference cost?” Now, a solo developer can burn through millions of tokens testing an idea and barely notice the bill. And that's the part of the AI story that doesn't get enough attention. Everyone talks about model intelligence. But AI pricing determines who gets to build. I've watched founders kill features because the token economics simply didn't make sense. If inference keeps getting cheaper this quickly, we need to ask a bigger question: What happens to companies whose entire business model is basically “we wrap GPT and add a markup”? Every major price reduction makes that model harder to defend. Cheap intelligence isn't just making AI more accessible. It's going to challenge a lot of the assumptions companies have built their products, pricing and roadmaps around. The cost of intelligence is collapsing. And that could be one of the biggest stories in AI. What do you think happens to AI startups when the underlying intelligence becomes almost ridiculously cheap? #AI #OpenAI #ArtificialIntelligence #Startups #ProductDevelopment #TechPricing #AIStartup

Jeff Bezos finally said what Prometheus is actually building: not your typical AI startup. The company raised $12 billion at a $41 billion valuation. The idea is what they call an "artificial general engineer," software meant to design and help build physical things: jet engines, spacecraft, chips, cars, drug compounds. His partner is Vik Bajaj, a physicist and chemist out of Google X (Wing, Waymo), later a biotech incubator co-founder. That background tells us this isn't a chatbot with an evocative name. Bezos was clear about why Claude and ChatGPT can't do what Prometheus promises to do. They were trained on text, code, images, mostly scraped off the internet. They're good at language. They know nothing about material science, thermodynamics, or how a supply chain actually behaves under stress because none of that was ever in the training data. So Prometheus has to build entirely new datasets from scratch, pulled from physical engineering itself. That's a different, much slower problem than anything an LLM has cracked so far. Bezos also made a much bigger claim: that this is a bet on invention itself, the same force behind the plow and the steam engine. Faster invention, he says, means labor scarcity, not job losses. Enough productivity gain that some households might not need two incomes anymore. Big claim. Worth chewing on with even if it makes us raise an eyebrow. Another thing that is different is that software moves fast. Code ships, gets tested, gets rewritten in days. Physical engineering doesn't work that way. You still need physical prototypes, physical certification, physical testing. No amount of AI design work skips that step. Incidentally, that's probably why Prometheus is also reportedly raising tens of billions for a separate holding company to buy the industrial businesses it wants to disrupt. Manufacturers don't hand over the operational detail an AI system needs to actually learn from, and owning the business is one way around that. Bezos is careful to call this early. About 150 people. No timeline. Many miles to go, in his words. So, does this work? I don't know. But it's a genuinely different bet than the one everyone else in AI is making right now. Hence it's worth watching closely.

Great insight. If you haven't had a look, The Entrepreneur Lab is a great resource. I had the privilege to test it as part of the City of Melbourne and Catalysr - Accelerator for Migrants & Refugees Migrapreneur Program.