Meta's Muse Turns Chats Into Project Boards That Update in the Background

Welcome to the Superpower Daily AI Briefing I'm super excited you're joining us for today's explainer We are unpacking four massive stories that are completely reshaping the future of artificial intelligence right now From the software organizing our daily lives to the literal final frontier of data centers Okay let's dive into this On today's agenda we're covering first meta adding persistent spaces to its muse AI agent Second open AI putting its newest models through a grueling real world

biology test Third a stunning claim that an AI has actually proved an 18 year old quantum computing conjecture And finally Elon Musk's wild timeline for putting SpaceX AI computers in orbit Meta spaces and muse AI organizing ongoing projects in the background So how exactly does this work It's a huge shift from the cluttered chat interfaces we're used to Step one you just send a normal chat request Let's say you're planning a trip Step two instead of

just spitting out a text answer muse actually builds a whole planner tracker or project board for you But step three is the real game changer Any connected apps you have will automatically update this workspace right in the background Meaning if your flight gets delayed or a calendar event shifts the board literally refreshes itself You don't even have to prompt it for new info It just becomes this living breathing project hub To put that into perspective Mark

Zuckerberg himself actually posted about this on X He said I built a barbecue game an event tracker and a home control dashboard Pretty wild range right From tracking ribs on the grill to managing smart home stuff It's a fascinating look at the versatility Meta is aiming for here But you know keep in mind these are his own personal examples not exactly an independent review of how it works out in the wild And speaking of the real

world we've got to separate the hype from the reality here Meta mentioned on their October earnings call that Muse hit 10 million users across the U S and Canada in just under two months Huge number absolutely But that 10 million is just the audience size for Muse as a whole It doesn't actually mean 10 million people are using these new spaces The real test is gonna be whether those background updates are reliable enough to keep us

coming back to these project boards day after day All right chapter two OpenAI evaluates lab reasoning BioBasicsBench tests wet lab science The big headline number here 66.7% OpenAI partnered with a biotech company called LIGO to create this massive 100 task benchmark to see how AI handles actual messy wet lab biology And when pushing GPT 6 Astra at its absolute highest reasoning effort it earned 66.7% of the available credit To give you some context on that GPT

5.6 Sol scored 43.8% on those exact same tests So we're talking about a really massive leap in capability But why is this specific benchmark such a big deal Well just look at the comparison Normally AI models are evaluated using beautifully polished scientific figures You know the perfectly curated charts you'd see in a college textbook But everyday science it doesn't look like that It is incredibly messy BioBasicsBench forces the AI to look at raw lab data We're

talking smudgy gel images imperfect microscopy raw videos of benchmark It has to troubleshoot conflicting observations on the fly just like a real human scientist has to do every single day That being said definitely don't hand the lab keys over to the AI just yet We need a quick reality check on some critical limitations First off that 66.7% that's a rubric score for partial credit The AI got points for its measurements and logic but it doesn't mean

it flawlessly nailed 66% of the experiments In fact Astra made a massive mistake on a cloning task It spotted the visual differences perfectly but then completely reversed what they meant misreading the control If human researchers followed its advice they'd have purified the completely wrong plasmid which by the way is a DNA molecule used in these experiments It also made some really basic visual errors like mistaking empty agar for biological colonies and over counting by about 12%

Bottom line it can see the details sure but it doesn't always reach the correct scientific conclusion AI proves quantum computing conjecture a breakthrough from physical superintelligence or PSI Okay this one is a bit of a mind bender To understand what they actually proved we need to talk about query cost It's basically just the number of times a computer algorithm has to read bits from its input data A normal classical computer reads one bit at a time

but a quantum computer can read multiple positions all at once Using something called superposition The 18 year old math problem they tackled the Aaronson Ambinus conjecture essentially says that for standard random inputs classical computers can approximate a quantum computer's answer with only a polynomial increase in reads not an exponential one So it basically puts a hard ceiling on the gap between quantum and classical computers for these kinds of unstructured problems But how this actually went down

is literally the coolest part PSI claims they ran instances of an unreleased anthropic model on their own internal infrastructure And by day three the AI spit out a complete mathematical proof Day three And this isn't just empty hype because they ran this proof through Lean Lean is a super rigorous proof assistant software that checks mathematical logic step by step And Lean totally accepted it Following that on November 6th PSI uploaded the full formal proof straight to

GitHub So instead of just dropping a bold press release they actually gave the math community a concrete artifact to inspect But and this is a big but we've gotta pump the brakes just a second Yes the Lean software verified the logical steps But as of right now this paper has not passed human peer review Independent experts still need to heavily scrutinize this to confirm what the theorem actually implies and to make absolutely sure the AI's setup

of the problem was flawless to begin with SpaceX Orbital AI Computers Elon Musk's star mined proposal For our last story today we're literally looking up to the stars SpaceX wants to put AI data centers into low Earth orbit in a massive project known as StarMind The sheer scale here is absolutely staggering Reportedly SpaceX is co designing the very first systems with NVIDIA using the Vera Rubin NVL72 platform as a starting point To give you an idea

the Earth bound version of that setup packs 72 Rubin GPUs and 36 Vera CPUs Now we don't know the exact payload specs for space yet but current projections point to a first launch in either Q3 or Q4 of 2027 with deployments heavily accelerating through 2028 and 2029 Elon Musk has been extremely vocal about this whole endeavor On November 6th he posted when we unveil the details of the StarMind computer next year it will blow people's minds

But listen to his phrasing carefully He's promising to reveal details in 2027 That's a commitment to share more info not a guarantee that a massive orbital data center network will actually be up and running by then His main economic argument here is that the abundant solar power up there could soon make space the absolute cheapest place to run AI Of course this ambitious plan faces some totally brutal engineering roadblocks Even Microsoft president Brad Smith has openly

questioned if this makes any sense at all Space is an incredibly harsh environment for data centers First you have to somehow remove the massive amount of heat these chips generate but in a vacuum where there's zero air to carry it away You have to shield hypersensitive electronics from constant cosmic radiation You have to figure out how to service the racks when a GPU inevitably fries And the whole shebang relies entirely on SpaceX's Starship achieving a completely

flawless rapid fire launch schedule just to get this heavy infrastructure into orbit And this brilliantly illustrates our key points Chapter five takeaways and final question Today's AI briefing summary Let's pull it all together Okay here's what you really need to remember from today First Meta's new Spaces aims to automate your projects in the background Though it's 10 million user audience doesn't yet prove real adoption of these persistent workspaces Second OpenAI's Astra is getting way better at

messy lab logic but a critical failure on biology controls proves we aren't at full AI lab autonomy quite yet Third PSI actually produced a software verified proof for a major quantum conjecture but the ultimate test of human peer review is still pending And fourth SpaceX is pushing forward with NVIDIA powered orbital AI data centers but they still have to beat the unforgiving physics of space and insane launch requirements Does the promise of incredibly cheap limitless solar

computing actually justify the mind boggling expense of launching AI data centers into orbit Or are the practical hurdles like heat radiation and maintenance just too high to ever overcome I'd love to hear your thoughts on this Read more at superpowereddaily com Stay curious