The Digital Tycoon
Meta gave away a model you can run on a desk, and the bond market rallied straight through the hottest wholesale inflation in four years.
A relief day, and the bond market did the relieving: the 10-year yield fell a tenth of a point, and ether bounced harder than bitcoin. The Nasdaq barely moved, a sign the AI selloff paused rather than ended.
Meta just handed builders a top AI model they can own instead of rent
On Thursday, Meta released Muse Spark Open, an AI model anyone can download, run and change. Meta calls it the most capable open-weight model from an American lab, built for coding and agents. It's also a U-turn: when Muse Spark launched in April, Meta kept it closed.
"Open weights" means you get the model itself, not a login to someone else's server. This one is huge, about 816 billion parameters, the dials it learned in training. But it only wakes up about 42 billion for each word it writes, which makes it practical. Meta worked with Apple so it runs on two linked Mac Studios, and it's live on hosts like Groq and OpenRouter.
Meta isn't giving away its best. Chief AI officer Alexandr Wang said Meta "will generally not open-source our most powerful models," and its top model, Muse Wonder, stays closed. The open one is the free sample: start on it locally, and when you outgrow the desk, Meta wants to be the cloud you pay. The other reason is China. "Most of the world's AI will run on open models. They should be American," Wang said. Alibaba and DeepSeek have owned the open shelf, and still lead on some coding tests.
The money angle is a squeeze on the middle. When a strong model is free, it's harder to charge a premium for a similar one by the word. The license spares small builders and catches giants: companies above $20 million in yearly revenue need a separate deal to host it as a service.
Why it mattersA model near the top of the field can now live on hardware you control, which means the price of AI isn't set only by the companies that rent it out.
If you pay an AI bill every month, you now have a real second option for coding and agent work. Test it on your own tasks through a cheap host before buying hardware. Read the license: products need a "Built with Muse Spark Open" label. And a model you run yourself can't be repriced or switched off by a vendor overnight.
A framework, not a call: when software gets cheaper, value tends to move to what it runs on. Watch who earns from open models (chip and memory makers, hosts like Groq and Nebius, Apple's high-end desktops) as much as who sells access by the word. Free rivals usually show up in the closed labs' margins before the headlines.
Sources · SiliconANGLE · Meta releases 816B-parameter Muse Spark Open · TechCrunch · Meta launches an open-weight model that runs on Macs · Hugging Face · Welcome Muse Spark Open · Briefs · Meta shares Muse Spark Open weights, keeps top-tier AI closed · The Next Web · Muse Spark launched closed in April
- CHIPS + ENERGYApple's CEO called the memory shortage an "absolute crisis" and won't rule out Chinese chips
In an interview with Time, Apple CEO John Ternus said "the constraint in memory is an absolute crisis" and that every option is on the table. That includes CXMT, a Chinese memory maker the Pentagon has blacklisted, whose chips Apple is reportedly testing. Its usual suppliers, Samsung, SK Hynix and Micron, can't keep up with AI demand, and Apple has already raised MacBook and iPad prices. When the richest buyer in hardware is short, everyone smaller pays more.
Go deeper · Seoul Economic Daily · Apple CEO signals openness to Chinese memory chips
- CRYPTOEurope put USDC under direct watch, the first stablecoin big enough to earn it
The European Banking Authority named Circle's USDC the EU's first "significant" e-money token under MiCA, Europe's crypto rulebook. It's the first of 45 such tokens to cross that line. Starting November 12, the EBA itself will supervise USDC's reserves, cash buffers and redemption plans, and it signed an information-sharing deal with New York's regulator. USDC's supply is about $73 billion. Closer watch is a cost, but it's also the badge banks and payment firms look for.
Go deeper · The Block
- BITCOINBitcoin slipped under $80,000 for the first time in a month, then bounced within hours
Bitcoin dipped below $80,000 on Thursday, its first time there in a month, as the selloff in debt-heavy AI stocks spread to anything risky. It climbed back about 3% within 90 minutes as Treasury yields dropped. Ether fell to a two-month low near $2,370 before rebounding. Neither move had much to do with crypto itself. For now, bitcoin is trading like a tech stock with a rate problem, and the bond market is holding the remote.
Go deeper · CoinDesk · Bitcoin bounces back to $82,000 after first fall below $80,000 in one month
- DIGITAL BUSINESSGoogle's Universal Cart is live in the US, and its checkout is built for AI agents
Google launched Universal Cart for US users in Search and the Gemini app. Shoppers save products from different stores into one cart, and Google watches for price drops, restocks and promo codes, even suggesting which card earns the best perks. Checkout runs on Google's Universal Commerce Protocol, an open standard that lets AI agents and stores talk directly. Amazon and OpenAI are building the same thing. More and more, the cart lives at the AI, not on your site.
Go deeper · TechCrunch
- DIGITAL MEDIATikTok signed a second studio to make creator-led microdramas
TikTok struck a multiyear deal with Made By Us Studios to co-produce vertical series for TikTok and PineDrama, its microdrama app. The first stars creator Hannah Stocking, who also executive produces. It follows a deal with Issa Rae's HOORAE, whose show Screen Time has drawn over 580 million views. HOORAE's next series packs 68 episodes into about an hour, with brands like Poppi written into the plot. Creators aren't just the cast anymore. They're getting producer credits.
Go deeper · Tubefilter · TikTok says yes to more creator-led microdramas
- ROBOTS + FUTURE TECHAmazon's Bee launched a $75 wearable that listens to your day and writes the to-do list
Bee, the AI wearable startup Amazon bought last year, launched two devices. The Frontier costs $74.99, clips to your wrist or shirt, and runs up to seven days per charge. The Eclipse is a card-sized recorder that snaps onto the back of a phone. Both turn conversations into summaries and to-dos, and the paid tier links them to Alexa+, Gmail and Outlook. Bee says audio is deleted after processing. Amazon's bet is that memory, not a screen, is the next interface.
Go deeper · TechCrunch
Thursday's producer price index, which tracks what businesses get paid for what they sell, rose 1.3% in September. That's more than double the forecast, and the biggest one-month jump since March 2022. Goods prices alone rose 2.6%, the most since the series began in 2009, with gasoline doing the heavy lifting. On a normal day, a print like that pushes bond yields up.
Instead, the 10-year Treasury yield fell about a tenth of a point, to 5.09%. Two things beat the inflation data. Money left AI stocks for the safety of bonds, and Fed officials talked down an October hike. "Why force urgency when you don't need to have urgency?" San Francisco Fed President Mary Daly told Reuters. Market odds of an October hike fell from 54% to 32%, though a December hike still looks likely to traders.
Why watch itWhen bonds shrug off the hottest wholesale print in four years, fear of an AI unwind is outweighing fear of inflation, and that balance sets the price of money for builders and holders alike.
Sources · TradingKey · US September PPI rises by most in over four years · Mecklenburg Times (Reuters) · Fed's Daly says rate hikes likely needed, but not urgently · Blockhead · Bitcoin holds near $82,000 as Fed officials cool October hike bets
A free open model is only a bargain if it can do your actual work, and most builders switch on gut feel. Instead, spend ten minutes building a small exam from your real tasks, then run any two models through it side by side. Open Gemini, paste this, and fill in the brackets.
I run [describe the business, e.g. a SaaS that writes product descriptions for Shopify stores] and I use an AI model through an API for [the jobs, e.g. writing descriptions, tagging products, answering support emails]. Build me a 20-item test set to decide whether a cheaper or open-weight model can replace my current one. For each item, give me: the exact input I'd send, what a great answer must include, what counts as a fail, and a 1 to 5 scoring rule. Split the items into 5 easy, 10 typical and 5 nasty edge cases. Then give me a scorecard table I can fill in for two models side by side, and tell me the minimum score that would justify switching, given that a bad answer costs me [e.g. a refund or an angry customer].
You'll get a reusable exam built from your own work. Run your current model and an open one on a host like OpenRouter through it, and by the afternoon you'll know whether switching saves money or just moves the problem.
Q. What's an open-weight AI model, and why would a company give one away?
An AI model is a giant file of numbers, called weights, learned during training. The weights are the model. With a closed model like ChatGPT, that file stays on the company's servers, and you pay to send it questions. With an open-weight model, the company publishes the file. You can download it, run it on your own machines, and retrain it on your own data.
Open weights aren't quite open source. You usually don't get the training data or the recipe, and the license can carry rules. Meta's new Muse Spark Open is free for most builders, but companies with more than $20 million in yearly revenue need a separate deal to host it as a service.
So why give it away? Three reasons. It's a funnel: developers who learn on your model tend to stay, and Meta wants them renting its cloud once their projects outgrow a desk. It's a price weapon: a free model makes it harder for rivals to charge a premium. And it's flag planting: whoever's model the world builds on gets to set the defaults. That's why Meta keeps its very best model closed. The free one is the front door, not the whole house.
Sources · SiliconANGLE · Muse Spark Open license terms · Briefs · Why Meta keeps its top models closed
| Fri 10/16 | Fed Chair Warsh talks AI and energy shocks with the IMF's Georgieva · The last senior Fed remarks before the pre-meeting blackout |
| Mon 10/19 | Comments close on Treasury's stablecoin issuer rule · Who gets to issue a dollar stablecoin in the US |
| Tue 10/20 | Comments close on the SEC's Regulation Crypto Assets · New exemptions for raising money with tokens |
| Tue 10/27 | The Fed's two-day meeting begins · October hike odds fell to about 32% after Daly |
| Thu 11/12 | The EBA takes over USDC's reserve oversight · Europe's first direct watch on a stablecoin |
Free models, scarce memory, and a bond market that picked its fear. Enjoy the weekend. MRTY