A model with 816 billion parameters, given away for anyone to run. Apple putting AI help in the kitchen. Wall Street shaking over the cost of the AI buildout. It is the Friday roundup on the Web Guy Nick Podcast, and Alex, Jordan, and Nick sort the week into what matters for a small business: the right size model for one job, short contracts while prices fight downward, and plain talk on what any of it means for your monthly tools. Listen on Spotify, and visit webguynick.com for AI help that fits a real business budget.
Listen to this episode on Spotify
That is the real shift for a small business. You do not need the biggest model. You need the right size model for one job, ideally one that can run close to your data.
Nick
You rent the giant. You do not house it.
Nick
I would not pause useful AI work because hedge funds had a bad week. A chip stock and a tool that saves your office manager five hours a week are two different things. Keep contracts short, keep your data exportable, and let the big players fight over price. Their fight is your discount.
Nick
Full transcript
Alex: Before we start: this podcast is produced with AI. Jordan and I are AI hosts, and Nick’s voice is an AI clone of his real voice. Every story is researched from public sources, with links in the show notes.
Alex: Here is the strangest fact of the week. Meta just gave away an AI model so large that it needs two desktop computers and about a terabyte of memory to run at home, and on the same week, the stock market punished almost everyone betting on AI. Giving it away and losing money, all at once. That is This Week in AI.
Jordan: I am Jordan, with Alex, and this is our Friday roundup. Five big stories, one item most people missed, and what to watch next week. Meta went open, Apple put AI on a kitchen screen, Google turned phone owners into app makers, Nvidia joined the open model flood, and Wall Street had its roughest AI week in months.
Nick: And I am Nick. I run a web design and marketing business, so my question on every story is the same. What does this change for a real business on Monday morning? This week, quite a lot, and most of it is good news for the little guy.
Alex: Story one is the model story of the week. On Monday, Meta launched Muse Wonder, its new frontier model. Then on Wednesday, October fifteenth, Meta released Muse Spark Open, an open weight model that anyone can download.
Jordan: The numbers are wild. Eight hundred sixteen billion total parameters, with about forty two billion active for each piece of text it produces. A context window of one million tokens, which means it can hold a huge pile of documents in one conversation. And it takes in text, images, audio, and video.
Alex: Meta and Apple ran it on two Mac Studio computers with the new M five Ultra chip, each with five hundred twelve gigabytes of memory, linked into roughly one terabyte. Apple testing showed about twenty five tokens per second on two machines. The smaller memory version of that computer already costs at least ten thousand seven hundred ninety nine dollars.
Nick: Let me save every owner listening some money. You do not need to buy the computers. That demo is a proof of concept, not a shopping list. The practical route is the hosted version. Together AI lists it at one dollar and twenty five cents per million input tokens, and four dollars and twenty five cents per million output tokens. You rent the giant. You do not house it.
Jordan: The scoreboard is honest, and Meta published it. On a tool use test called MCP Atlas, Spark Open scored eighty nine point two, ahead of the Chinese open models compared. But on a coding test, Terminal Bench, it scored seventy six point four, behind Kimi K three at eighty five. Best at using tools, not best at writing code.
Alex: The license matters too. Products built on it must say Built with Muse Spark Open. A company offering it as a hosted service, once yearly revenue passes twenty million dollars, needs a separate agreement with Meta first. Internal use is exempt.
Nick: Here is the owner angle. A strong open model means the price of good AI keeps falling, and you get leverage. A model you can run inside your own walls, or rent from a smaller host, is a real plan B. Plan B is what keeps a small business from being trapped when a vendor raises prices.
Alex: Story two is Apple, and it is a different kind of AI launch. On Tuesday, October thirteenth, Apple unveiled HomeView, a smart home display built around its new Siri AI assistant.
Jordan: It is a seven inch screen for the counter or the wall. Both versions cost two hundred ninety nine dollars, and it ships on November thirteenth. It recognizes up to six people by face, showing each person their own calendar, alerts, and screens. One limit. It will not be sold in the European Union at launch, because the new Siri AI features are not switched on there.
Nick: Owners should look past the gadget. The AI assistant is leaving the phone and moving onto a shared screen, where families decide what to buy and who to call. If your business shows up correctly when a kitchen screen is asked for a recommendation, that is the new front door. So make sure your hours, services, and service area are accurate everywhere, because a shared screen reads them out loud to the whole room.
Alex: Story three might be the most useful one for owners, and Google announced it the same day. Google Play now lets a regular person build a custom Android app with AI and run it on their own phone, with no developer account.
Jordan: Here is how it works. You describe the app in Google AI Studio, using plain words. The AI writes a real Android app, and Google Play installs it on your phone as a personal app. It stays private, never in the public store or in search. The rollout started in the United States for adults, and sharing with a small circle is coming in the next few months.
Nick: I love this one, because every owner has little tools they wish existed. A job site checklist. A quote calculator for one service. Before this week, each was a developer project. Now the first version is a description you type on a Tuesday night, tested on your own phone on Wednesday. Keep it private, and keep customer data out of it while you learn.
Alex: Story four is the flood. Meta was not alone. On Thursday, October fifteenth, Nvidia released Nemotron three point five Super VL, an open model with one hundred twenty billion total parameters and about twelve billion active, built for enterprise agents that read documents, images, and video.
Jordan: Nvidia also released smaller Bolt models for local machines. One fits a seventy four billion parameter model into twenty four gigabytes of memory, small enough for one strong workstation. A smaller four gigabyte version arrives on October twenty third. Companies can train Super VL on their own data, and Nvidia named Atlassian, CrowdStrike, and Salesforce among those integrating it.
Nick: That is the real shift for a small business. You do not need the biggest model. You need the right size model for one job, ideally one that can run close to your data. A model on a workstation in your office reading invoices is a different conversation from sending every client file to a cloud chatbot. Efficiency is becoming the product, and efficiency lowers your bill.
Alex: Story five is the money, and it explains the nervous week on Wall Street. Reuters reported this morning, citing a Goldman Sachs note, that global stock picking hedge funds lost money every single day this week, down two point nine percent for the week through October fifteenth.
Jordan: Asia took the hardest hit. South Korea’s main index fell eight percent on Thursday to a six month low, and Taiwan stocks lost more than three percent. The funds are still up fourteen point three percent for the year, so this is a shake, not a collapse.
Alex: The worry underneath is familiar. Investors are asking whether all this AI spending pays off soon enough, especially with borrowing costs staying high. It connects to the warning we covered Wednesday, when Gartner said revenue to cover the data center buildout may fall trillions short.
Nick: I would not pause useful AI work because hedge funds had a bad week. A chip stock and a tool that saves your office manager five hours a week are two different things. Keep contracts short, keep your data exportable, and let the big players fight over price. Their fight is your discount.
Alex: A quick recap of the rules story, because it shaped the week too. Thursday we covered in depth how eight states sued OpenAI over ChatGPT and children, and how Europe sent its first AI Act information requests to OpenAI and Anthropic. No fines, no rulings yet. Watch for OpenAI’s first responses, because eight separate courts means eight timelines.
Jordan: Time for Higgsfield Corner. This week the useful update is about scheduling. A Higgsfield guide published on October twelfth lays out how creators can request batches of video through a chat assistant, or schedule recurring jobs with its Supercomputer workspace.
Nick: Scheduling sounds small until you run content every week. Product clips rendering overnight beats babysitting generations at your desk. Two cautions. Every scheduled run still costs credits, so a recurring job can spend quietly while you sleep. And a human still has to watch the finished video before it posts. Schedule the production, never the publishing.
Alex: Now our quick fire game, AI or BS. Three statements. Nick, you are playing. Statement one. Meta’s new open model can run on two Apple desktop computers with about a terabyte of combined memory.
Nick: Real. Wild, but real. Two Mac Studios, five hundred twelve gigabytes each. I still say rent it instead.
Jordan: Correct. Statement two. Apple will sell HomeView in the European Union on launch day, November thirteenth.
Nick: That is BS. Apple is skipping the European Union at launch because the new Siri AI is not on there. The date is right, the place is wrong.
Alex: Correct. Statement three. Hedge funds that pick stocks lost money every day this week, according to a Goldman Sachs note reported by Reuters.
Nick: Real, and down two point nine percent for the week. A bad week, but they are still up for the year, which is exactly why owners should not panic over headlines.
Jordan: Three for three again. Now the item most people missed. On October thirteenth, a company called Liquid AI released a model with twenty four billion total parameters and only two billion active at a time.
Alex: It is called LFM two point five. In testing tracked by BenchLM, it beat a larger Alibaba model on tool use, while losing on a standard coding test. Small, fast, and built to be good at one job instead of everything.
Nick: That little model is the whole week in miniature. Two billion active parameters doing useful agent work on modest hardware. For an owner, tools that run on a phone or a shop computer keep getting better and cheaper to run, even while the giants fight at the top.
Alex: Which brings us to one thing to try this week, and it uses story three. This weekend, build one private personal app for your business in Google AI Studio.
Jordan: Pick one small internal job. A daily opening checklist, or a price estimator for one service. Describe it, let the AI build it, and install it on your own phone through Google Play. Nobody sees it but you.
Nick: Give it thirty minutes and judge it honestly. If it saves you one paper form, it earned its keep. If it flops, it cost you half an hour, not a developer invoice. Try small, on your own phone, with your own real job.
Alex: What to watch next week. Independent test results for Meta’s open model, now that everyone can download it. The smaller Nvidia Bolt model arriving on October twenty third. Whether the stock selloff calms down or spreads. And the first responses in those eight state cases against OpenAI.
Jordan: That is the show. A giant model given away, AI on a kitchen screen, an app anyone can build, and a market reminder to keep your contracts short and your data yours.
Nick: Do the one thing. Build that one little app this weekend, and test it on a real job on Monday. If you want help making your website ready for the AI screens your customers are about to talk to, you will find us at webguynick.com. See you Monday.
