The Signal / Superpower Daily
Anthropic weighs a model as Astra gains ground
Competition, deployment controls, and real-world testing shape the AI radar today. Anthropic is weighing a response to Astra’s commercial momentum, while new research and reported incidents put more pressure on how systems are evaluated before they act.
Superpower Daily: The Signal
Listen to this episode
Episode guide
Show notes
Competition, deployment controls, and real-world testing shape the AI radar today. Anthropic is weighing a response to Astra’s commercial momentum, while new research and reported incidents put more pressure on how systems are evaluated before they act.
In this episode
Full transcript
Read along
Select any transcript timestamp to continue listening from that point.
Welcome to the signal from superpower daily with Maya and Theo We have a massive amount of ground to cover today Yeah we really do So we're gonna start right away with anthropic because they are currently weighing a brand new model release Right as open AR as GPT 6 Astra is really gaining ground right The pressure is absolutely mounting over there It really is So let us jump straight into this lead story Anthropic is internally debating this new release and you know there is no official launch commitment yet No they're just actively deliberating but the timing of this deliberation is what is so critical I mean
it is happening right on the heels of open AI launching Astra which was September 3rd right Exactly Yeah September 3rd and you have to look at the exact phrasing open AI used for that launch Because they describe Astra as their most capable broadly deployed model broadly deployed Yeah Yeah that is very specific phrase It is deliberate It signals a major shift in their strategy They are aiming directly at total enterprise dominance right now And so the context around anthropics internal debate is just fascinating to me Because it is driven entirely by this competitive and financial pressure You are seeing this pressure mounting on like multiple
fronts all at once We really need to look at the actual enterprise spending data here because this is where the narrative hits actual reality right We have data from ramp which for you listening ramp is a platform that tracks corporate card and invoice spending It is incredibly valuable information really is it shows us where actual businesses are swiping their credit cards today Yeah it tracks deployed capital It does not track theoretical interest or You know marketing buzz it tracks real corporate budgets and the ramp data is very clear It shows Astra currently commands about 13 of tracked enterprise AI spending 13 is a significant chunk
of the market right out of the gate It is meanwhile Anthropics flagship model which is Claude fable is currently sitting at about 8 of that track spending You do have to frame those numbers carefully though I mean that is a specific slice of monitored spending on the ramp platform right Right It is not a perfect measure of the entire Global AI market Yeah because some massive enterprise deals happen completely off platform But it is still a very telling slice It shows the velocity of Astra adoption definitely and then you look at the developer side Because we have new data from open router open router is
an aggregator platform Yeah It is basically a place where developers go to access many different AI models through a single interface And they reported a massive shift in usage for the first time in over two and a half years Their users actually spent more on open AI models than on Anthropic models two and a half years is an absolute eternity in the AI industry It really is an eternity and prop deck has historically dominated developer preference on that specific platform right Developers often preferred Claude for coding or complex reasoning tasks So losing that specific lead is jarring it perfectly explains why a new model release
is suddenly under active consideration Okay I want to push back on this panic narrative just a little bit Oh really Yeah I mean we are looking at a 13 to 8 gap in the ramp data That is literally a 5 difference in a narrow slice of enterprise spending You think 5 is too small to trigger a panic inside a company Well exactly imagine you are a pre IPO company You have a long term roadmap is a 5 market share dip Really enough to force you to rush a frontier model out the door Or are we just looking at standard scenario planning here I mean every
smart company runs contingencies a Contingency plan leaking to the press does not mean the sky is actually falling you are right to question the scale of the panic But it is not just about the static 5 difference It is about the trajectory You mean the momentum going into their initial public offering exactly an IPO is completely built on a growth narrative Yeah you have to convince Wall Street that you were capturing the future right If your momentum stalls right before you ring the bell your valuation just tanks But here is where the situation gets incredibly complex because the economics of deploying a frontier model right
before an IPO are Absolutely brutal We need to unpack those economics because if deploying a new model wins back market share Why is it a brutal decision Well you are caught in a fundamental tension You basically have two conflicting goals You need to show market dominance You need to prove you have the absolute best technology right But your second goal is proving a clear path to profitability Wall Street wants to see your margins actually expanding and running a brand new frontier model completely destroys your margin exactly I mean training a model cost tens of millions of dollars but that is just the start The real
cost is inference right Inference is the cost of actually running the model every single time a user sends a prompt it costs an absolute fortune to serve these Massive cutting edge models to millions of enterprise users The computing power required is just staggering you are burning through server capacity and Investors are getting very impatient right now They're starting to ask very hard questions They want to know when these AI companies will actually turn a profit So you have a terrible choice to make you can launch a new model to beat Astra You might win back that 5 market share but doing so ruins your profit
margins right Analysts are digging through your financials or you know you hold off on the new model you protect your margins But you just sit back and watch Astra eat your market share That is the economic hurdle It is a massive dilemma But anthropic faces a second internal hurdle that is equally massive You are talking about safety Yes the Reuters report makes this constraint very clear Anthropic is evaluating if the prospective new model is actually safe enough to deploy Right and the safety review is not just a procedural detail for them It is the core of their entire brand identity They literally founded the company
on the premise of building safer AI they pioneered constitutional AI Anthropic positions itself as the responsible safety focused lab That is their unique selling proposition to the enterprise market Corporate clients buy Claude because they trust it will not hallucinate catastrophic legal advice They trust it will not generate toxic content exactly So if anthropic rushes a model out the door simply to answer a competitive challenge from open AI They risk that entire reputation right If the new model goes rogue or exhibits unpredictable behavior the brand damage is totally irreversible So they are actively weighing their technical readiness against immense commercial pressure and this internal debate reveals
something Fundamental about the whole industry right now the actual constraint on the AI race has shifted It used to be all about compute like the bottleneck was just hoarding enough GPUs to train the models that era is ending The bottleneck is no longer just securing compute The bottleneck is now margins and safety reviews exactly Can you afford the inference costs to run the model at scale and is the model Verifiable enough to release without destroying your company that completely reframes how we should view this arms race It is not just a sprint to build the biggest brain anymore No it is a delicate balancing act
It is about who can afford to deploy intelligence safely So the competitive pressure from Astra does not automatically settle the decision for anthropic right The evidence suggests this is still a highly contested internal choice They have not committed to a launch So the key thing to watch next is pretty straightforward Do they actually approve the release of this new model or Do they hold it back their final decision will reveal their true internal priorities if they hold the model back It sends a very clear signal It tells the market that their commitment to safety targets outweighed their fear of Astra or it tells the market
that their margin targets Were too fragile to absorb the inference costs true either way holding back proves They are not purely reactive to open AI until that choice is finalized The real story is just the immense pressure inside the race Yeah we are watching a pre IPO company try to balance urgency against astronomical costs and major safety risks It is a very delicate scale Mm hmm one wrong move could totally derail their public offering Okay that wraps up our look at the corporate AI race We are now shifting our focus entirely We are moving from base models to how AI is being applied to decode
physical reality We are looking at biology you see San Diego researchers just published two incredible studies both of these studies were published in the journal cell and The researchers essentially built virtual versions of living cells right They successfully paired artificial intelligence with advanced physics They focus specifically on mitochondria Mitochondria are crucial structures They help cells turn nutrients into usable energy people always call them the powerhouses of the cell They do it is a classic line But these two studies take entirely different approaches to building a virtual model the contrast between the two methods is what makes this research So important let us break down the
first study This team used a deep learning pattern finder and they called the system Mito space the scale of the training data for Mito space is staggering They trained this system on 30 000 microscopy movies and these were not standard flat images They were full 4d microscopy movies of cancer cells We need to explain that fourth dimension Yeah because they used a technique called lattice light sheet microscopy right This technology captures cellular structures in three full dimensions It uses ultra thin sheets of light It maps the physical space perfectly Yeah but the fourth dimension is time This is critical They are not looking at a
frozen snapshot of a cell They're actually watching the cellular networks move and change over time They took these living cancer cells Then they expose them to 25 different chemical compounds and these are drugs specifically known to disrupt mitochondrial function right Because mitochondria do not just sit still they form a highly interconnected network inside the cell that network is Constantly moving the mitochondria physically split apart which is called fission and they constantly fuse back together So a traditional flat two dimensional image completely misses all of that dynamic changing behavior exactly You cannot understand a traffic jam by looking at a single photograph You need a video
That is a great way to put it and the 4d data proved to be an absolute game changer for the AI The accuracy jump was phenomenal Mito space was tasked with classifying these 25 drugs It had to identify them by their mechanism of action And when the researchers restricted the AI to flat 2d images the accuracy was only 56 That is barely better than flipping a coin in some complex diagnostic scenarios But when they fed Mito space the full 4d movies the accuracy leaped to 75 That is a massive improvement The system basically learned how the mitochondrial networks physically move It learned the rhythm of
the fission and fusion It did not just memorize how things looked in a frozen moment and it achieved this accuracy without being given the drug labels during Its training phase right it learned the structural differences purely by observing the dynamic networks It learned to predict the entire Energetic state of the cell and it did this just by analyzing the shape and movement of the mitochondria It successfully tracked these changes across 26 different drug conditions and you know Here is where the research gets slightly mind blowing the Mito space system demonstrated zero shot capabilities right For those listening a zero shot capability means the AI can
successfully analyze something It was never explicitly trained to recognize it identified correct structural patterns in entirely new drugs These were compounds it had absolutely never seen during its training process It went even further than that It actually successfully sorted human lung organoid cells by their specific Developmental stage and it did that sorting without requiring any retraining of the underlying model which proves the model is actually learning something Fundamental about cellular biology Yes it is abstracting the actual rules of cellular energy It's not just memorizing the pistol patterns in the training data This whole process reminds me of weather forecasting Think about how meteorologists predict a
storm right These statistical models those models in just decades of historical weather data They hunt for recurring patterns Exactly Mito space operates exactly like a statistical weather model It finds hidden biological patterns and massive amounts of historical cellular data But modern weather forecasting does not rely on statistics alone It also uses atmospheric physics models Right those physics models are built on hard mathematical rules They use thermodynamics and fluid dynamics to basically simulate the atmosphere and that perfect analogy brings us to the second UC San Diego study Yes This second team built a physics based digital twin of a cell this companion study starts from a
completely different Philosophical premise they did not look for statistical patterns in massive data sets instead They mathematically modeled the actual physical machinery inside the cell they precisely mapped the mitochondria Then they mapped the microtubule tracks right Microtubules are basically the cellular highways the mitochondria travel along these tracks to get where they need to go The team mathematically modeled the transport system This movement is driven by specific motor proteins so they built a digital twin of a cancer cell based entirely on the laws of physics and they Carefully adjusted this digital simulation they tweaked the parameters until its baseline behavior Perfectly matched a real observed cell
under a microscope once the simulation matched reality They gave the digital twin a blind test They simulated the introduction of a specific drug treatment The drug was no cortisol in the physical world No cortisol Partially breaks down those microtubule tracks It destroys the cellular highways So the researchers ran the physical simulation of the drug interaction and crucially they did not tweak or change any underlying Model parameters to force a specific outcome They literally just let the physics engine run the digital twin successfully reproduced the correct biological response It accurately showed the reduced movement of the mitochondria It accurately showed the altered rates of fusion and
fission the simulation perfectly matched the physical behavior Observed in real living cells treated with no cortisol So returning to your weather forecasting analogy We now have two distinct tools We have a deep learning statistical model finding the dynamic patterns That is mito space and we have a deterministic physics based model proving the actual mechanical causes That is the digital twin The researchers have stated their ultimate goal here They eventually want to make these two distinct systems talk to each other They want to seamlessly combine the rapid pattern discovery of mito space with the rigorous physical simulation of the digital twin That unified system would be
incredibly powerful They want to add more cellular organelles beyond just mitochondria Eventually they want to scale this up to model entire complex tissues But we do need to establish a very firm caveat here because the validation of these virtual cells is still in its absolute Infancy right These systems are emerging as very powerful Exploratory tools they can heavily complement physical laboratory experiments but they absolutely do not replace physical lab work You cannot bypass biology No You cannot just run a clinical drug trial on a digital twin and skip the messy reality of the wet lab The physics engine is good but it is not a
perfect replica of a human body yet But the success with dynamic mitochondrial data proves a larger point It proves that virtual cell systems are fundamentally viable It really represents a massive foundational step forward for the future of computational drug discovery We are going to shift our context now We just explored how scientists are using digital twins to understand biology Now we are going to look at the highly controversial rise of digital twins in Hollywood the environment changes But the core theme of synthesizing reality remains exactly the same We need to examine the strange and disruptive case of Tilly Norwood Tilly Norwood is not a human
being Norwood is an entirely AI generated actor This system is a completely synthetic performer designed for feature films Norwood is currently starring in an upcoming feature film The movie is called misaligned the studio sent this synthetic system on a traditional press tour to promote the movie Which is wild and that PR tour hit a very bizarre and revealing snag on live television Norwood was booked for an appearance on Piers Morgan uncensored the veteran human actor Tom Conti was also a guest on the panel and Conti clearly had an agenda He pressed Norwood with a very direct uncomfortable question He wanted a clear answer about the
production of misaligned He asked if the cast of the film was human or computer generated The AI system attempted to deflect the question at first it relied on industry buzzwords It called the film a hybrid production right Norwood tried to emphasize the human element It pointed out that real human directors were involved It mentioned human writers and human editors it argued that human creatives were strictly running things behind the camera But Tom Conti refused to accept that evasive answer He repeated his original question forcefully He demanded to know if the other actors appearing on the screen were human or not Norwood was completely cornered Finally
admitted the truth The other on screen performers were all digital twins They were entirely synthetic creations exactly like Norwood itself and right at that moment of admission The system suffered a massive glitch Norwood was in the middle of explaining the technical definition of a digital twin Then it abruptly stopped speaking English The AI switched entirely into fluent Cantonese It was a highly visible live malfunction on national television It completely derailed the interview Norwood eventually recovered and returned to English the system immediately attempted to apologize It casually referred to the incident as a little hiccup and Piers Mergen pressed the AI He asked what exactly had
just happened to its programming Norwood offered a very vague excuse It claimed its wires just get a bit crossed sometimes but the PR disaster continued later that day Norwood gave a follow up interview to the News Nation Network During that second interview the AI offered a completely different excuse for the Cantonese incident It rebranded the malfunction It called the glitch a spontaneous linguistic flourish The system claimed that obscure conceptual connections simply surface unexpectedly when it is speaking Neither of those explanations provide an actual verifiable technical cause for the failure No they are essentially real time PR spin generated by a struggling algorithm Let us analyze
the technology here Why would an AI trained to speak English Suddenly switched to Cantonese under pressure large language models map concepts in incredibly high dimensional space Languages share overlapping conceptual neighborhoods in that space So when Conti aggressively interrupted the AI he disrupted its predictive token Generation exactly the system was forced to recalculate its response path instantly It likely hit a low probability token path that basically jumped the language barrier it is not a crossed wire it is a Fundamental architectural vulnerability and how neural networks process language and this live failure Exposes the massive irony of the entire synthetic actor enterprise Think about the core value
proposition being sold to Hollywood executives The studio wants a synthetic actor because they want total absolute control the studio executive wants an actor who never gets tired on a 14 hour Right they want an actor who never demands bigger trailer They want an actor who never goes off script during a crucial scene They are buying absolute predictability They want to eliminate human volatility from the production schedule So a spontaneous linguistic flourish is the absolute last thing a studio executive wants to see If your million dollar AI actor glitches and starts speaking Cantonese on Piers Morgan Your entire marketing strategy is hijacked you completely lose control
of the narrative the technical failure becomes the only event Anyone talks about the actual movie they were trying to promote becomes a complete afterthought But this disastrous interview exposed something much deeper about the state of the film industry in 2026 It exposed the true working definition of a hybrid film in this new era The industry definition of hybrid has entirely changed the human contribution is now strictly relegated to off screen roles The directors are human The writers are human The editors are human but the talent actually appearing on the screen is entirely digital It is no longer a mixed cast of humans and AI performing
scenes together It is a complete wholesale replacement of the on screen talent pool And this realization is pouring gasoline on the ongoing sag after a labor dispute The acting industry is already fighting a massive existential battle against the encroachment of synthetic performers Real human actors are absolutely furious about the normalization of digital twins Prominent actors have publicly condemned the talent agencies that are choosing to represent software like Norwood right Melissa Barrera Mara Wilson and Nicolas Alexander Chavez have all spoken out aggressively Mara Wilson asked a very pointed ethical question about the production of misaligned She asked why young human women were not hired to play
those roles instead that question cuts to the core of the labor issue It is especially potent considering how Norwood was created because Norwood's digital face is likely a mathematical composite It was generated by scraping the training data of thousands of real Uncompensated human women Nicolas Alexander Chavez was even more direct in his criticism He stated bluntly that Norwood is not an actress actually he is arguing that acting requires human lived experience Software cannot act it can only simulate emotional outputs There's a language switch on live television certainly does not settle that complex labor dispute No But the glitch highlights the immediate practical risk of relying
on these public tech demos the crucial question the industry has to watch next revolves around basic reliability How consistently can a synthetic actor actually explain its own film production without Hallucinating and can the underlying model avoid completely glitching on live television again because right now despite the Hollywood hype the technology looks incredibly fragile under pressure a Spontaneous AI glitch is embarrassing on a Hollywood PR tour But our next story demonstrates what happens when that exact same technology Hallucinates in a military context right we're moving to a newly reported incident and it shows how terrifying an AI Hallucination can be in the real world CNN just
reported that a chatbot assisted false assessment nearly led the United States military to board a Chinese ship This was a severe near miss scenario It risked sparking a major geopolitical escalation Let us carefully walk through exactly what happened in this incident United States military forces in the Middle East Actively prepared to board a Chinese commercial vessel They initiated this preparation based on a specific intelligence assessment and that assessment explicitly claimed the ship's cargo included Components intended for a nuclear weapons program according to multiple sources speaking to CNN The military response was immediate and serious military aircraft were scrambled and airborne Armed personnel were actively prepping
their gear for a hostile boarding operation Fortunately the interception never actually happened the chain of command halted the operation There was no armed confrontation at sea But the situation came dangerously close to a kinetic event one intelligence source stated clearly that this error Risked an outright escalation between the United States and China and that immense risk was generated based on entirely false Hallucinated intelligence the most critical part of this story is tracing the failure chain We have to understand how this false intelligence Successfully infiltrated the operational planning pipeline the failure started with a single human analyst this analyst worked at Special Operations Command Pacific the
analysts used an AI chat bot to examine the shipping manifest of the Chinese vessel The chat bot was fed a mix of different intelligence streams it received publicly available Open source material but it also received highly classified Signals intelligence it ingested intercepted communications the neural network mixed those complex inputs together It attempted to find patterns in the data And it reached a completely wrong conclusion about the nature of the cargo so the AI Basically hallucinated the existence of the nuclear components it connected dots that did not actually exist in reality Why does a model hallucinate nuclear weapons it predicts the next most likely word based
on its training weights right Military shipping manifests often share linguistic context with weapons reports in the models training data the AI just followed a Probabilistic path to the wrong word but we have to be clear about accountability here The software did not launch the military operation on its own No it is not a rogue AI launching missiles This is a story about a massive human workflow vulnerability the analyst read that hallucinated conclusion They did not independently verify the claim instead the analyst used the exact same AI tool to draft the formal written Intelligence report there was absolutely no Independent verification step between the chat bots
raw analysis and the generation of the formal document Which is wild because the standard advice in the tech industry is always to keep a human in the loop We hear that constantly But in this incident the human was explicitly in the loop the human was the one operating the software The human presence did not prevent the failure the human essentially rubber stamped the AI Hallucination the analyst took a statistical guess from a chat bot and gave the immense authority of an official military document that formatting is Incredibly dangerous once a hallucination is packaged to look like a formal intelligence report it completely bypasses critical thinking
it enters the operational Planning pipeline it has the correct headers It has the standard bullet points It looks authoritative so commanders and armed personnel made serious tactical decisions based on a flawed AI Generated document this incident clearly exposes how easily unverified AI text Can manipulate human hierarchies and this vulnerability is being exposed at a very precarious time for the United States military Right Defense Secretary Pete Hegseth launched the massive AI Acceleration strategy back in January that strategy is aggressively pushing AI models into the hands of three million military Personnel the deployment covers all branches and all classification levels you are rapidly scaling a technology that
has known fundamental hallucination issues to millions of individual users and the most alarming part is that the military branches currently lack a Unified common system for validating this AI generated intelligence the different branches use different software tools They have different operational rules They implement different safety checks this fractured approach basically guarantees that more hallucinations will slip through the cracks how does the loop itself need to be completely redesigned because Relying on a human just looking at a screen is clearly failing an AI system can be a highly useful tool for rapidly sorting Massive piles of information it can find needles in the haystack But it
remains fundamentally unsafe as the final source of a consequential judgment Turning an AI conclusion into a formal report makes an unverified claim travel much farther and much faster it gains Institutional momentum it becomes accepted truth simply because it is formatted correctly the missing safeguard is mandatory independent verification a Rigorous human led checking process must stand between an AI generated claim and the start of operational preparations What we need to watch next is the Pentagon's official response to this near miss right well the Defense Department Finally implement a mandatory shared verification requirement across all branches They absolutely have to mandate that commanders check AI generated intelligence
against secondary sources before operational planning begins You simply cannot have armed personnel prepping to board a foreign vessel based on a probabilistic chatbot error The geopolitical stakes are simply too high for this kind of algorithmic rubber stamping all right We are gonna move into our quick reads now We have three rapid fire updates tracking AI deployment across different sectors first up We have new results from the brood war bench testing suite AI agents can successfully win Starcraft matches But they cannot actually master long term strategy an AI agent called codex astro was tested at the highest difficulty settings It achieved a perfect 18 and 0
record on a Starcraft Brood war benchmark a perfect undefeated record always looks incredibly impressive on a press release it does But the creator of the benchmark Ben Swerdlow notes a very sharp caveat to these results Swerdlow points out that every single AI system tested is still playing at a fundamental beginner level These systems are not playing like human grandmasters They are relying on cheap tricks right the agents win their matches by ruthlessly exploiting AI hesitation They use a tactic called early worker harassment They send a basic probe unit across the map immediately the probe attacks the opponent's resource gatherers this forces the opposing AI to
pause and calculate a defensive response and that brief moment of Computational disruption is enough to win the early game but the victorious agents completely fail at sustained Complex economic management they completely fail at continuous unit production over a long match Swerdlow discovered a massive coordination problem inside the Architecture of these agents right the agents use separate sub networks to handle two different tasks one sub agent handles the economy Another handles the army but these sub agents barely communicate with each other during the game Military units are sent forward one at a time to die the AI cannot coordinate a massive synchronized push They can win
short matches through disruption but they entirely lack coherent long term strategic planning next up The robotics company figure evaluated its new humanoid robot in unseen environments figured tested the helix 2 5 humanoid robot Inside 30 different homes across the Bay Area and these specific homes were completely excluded from the robots training data It had never seen them before the robot achieved zero shot success on 56 of its full task trials the trials required the robot to tidy a messy living room fold towels and Successfully make a bed the robot collected absolutely no new data in those homes prior to the test It received no localized
fine tuning figure credits the success to their index pre training method Index is a massive data set composed entirely of human behavior video the robot basically watched humans do chores and Generalize the physical movements it is a highly notable attempt to measure how software skills transfer into real Unpredictable physical spaces but you do have to look closely at the grading system figure used for this benchmark They used a very strict pass or fail grading metric They did not award partial credit for incomplete tasks That means 44 of the assigned tasks were complete total failures The robot had to successfully perceive the room geometry walk to
objects without falling and use complex two handed manipulation So this test proves the transfer of very specific pre trained skills it is absolutely not proof of an open ended general household robot that can handle any random chore you ask of it a 56 success rate is real progress But dependable general household automation remains far from a solved problem Finally anthropic is aggressively moving into the realm of physical biology They just set up a new In house wet lab facility in the San Francisco Bay Area Eric Conner Abrams is leading this new life sciences Initiative for the company They are no longer just building models They're
actually running physical biological experiments themselves They want to experiment with allowing their Claude model to directly control robotic laboratory equipment They are testing autonomous science with very limited human intervention in the loop But anthropic is drawing very strict red lines around the activities permitted in this new fidelity The lab is not specifically designed for commercial drug discovery and they are strictly prohibiting the facility from running any form of clinical trials They are building hands on physical experience with AI driven robotics But human oversight is still deemed absolutely essential for maintaining safety protocols They're keeping autonomous lab operation firmly on the exploratory side of the safety
line for now that wraps up our quick reads for today We are moving to three takeaways from today first There is an escalating tension between deployment speed and operational safety across the entire AI ecosystem You see this tension clearly in anthropic debating a frontier model delay for internal safety reviews You see it violently in the near miss military boarding caused by unverified AI Intelligence the capabilities of these systems are simply moving faster than our institutional verification protocols can handle second We are witnessing the rapid bridging of the digital and physical worlds UC San Diego is building digital twin cells to simulate complex drug physics figure
is putting zero shot humanoid robots into physical living rooms and Anthropic is building physical wet labs to let Claude control robotic biology equipment The models are no longer confined to the screen third the illusion of human control is becoming a massive systemic liability AI agents can sweep a Starcraft leaderboard but completely fail to coordinate basic long term strategy a Synthetic actor glitches on live national television and totally hijacks its own press tour These synthetic systems remain highly unpredictable when they are forced to improvise outside their training bounds The crucial development you need to watch tomorrow is whether the Pentagon formally responds to the CNN report
Well the Defense Department finally institute new mandatory verification mandates for intelligence analysts using chatbots We really need to see if a shared independent checking protocol becomes standard military policy before another hallucination triggers a crisis That is the signal for today You can find more updates and our full archive at superpower daily com Thank you for listening We'll see you tomorrow
Original reporting
Stories covered
Read the complete Superpower Daily coverage behind this episode, including reporting context and source links.
