Apple alleges OpenAI used its circuit schematic

Welcome to Superpower Daily Today Apple drops new evidence in its trade secrets fight with open AI alleging an AI agent was used to tune stolen schematics Let's get into it Yeah this is a scenario that genuinely pushes the boundaries of what we traditionally consider corporate espionage If you look back at the history of Silicon Valley IP theft like the famous Waymo versus Uber self driving lawsuit it usually involves a human like someone downloading gigabytes of files on their way out the door Just to give their new human engineering team a headstart Exactly but what Apple is alleging here in this August 31st federal court filing

introduces a completely novel layer of automation to the concept of stolen intellectual property Right because they're claiming a former Apple engineer named Chang Liu didn't just walk away with the blueprints According to Apple's request for an injunction Liu downloaded a confidential circuit schematic in March Which crucially was two months after he actually left his job at Apple Yeah that's wild He then allegedly took this highly sensitive schematic and integrated it into his new hardware development work over at OpenAI And you know that two month gap is the first massive red flag In modern enterprise security the moment an employee gives notice their access is typically

monitored heavily Oh yeah And it is entirely revoked the second they're off boarded So for a former employee to allegedly reach back into the network and extract a confidential schematic 60 days post departure It's bad Yeah it suggests either a severe lapse in Apple's access protocols or a highly sophisticated circumvention of those systems But what happens after the alleged download is where the technical implications just get truly fascinating Okay look it up for us Apple claims that forensic analysis of a company MacBook that Liu retained shows he imported the schematic into a software environment known as LTSpice Okay so for those of you who haven't

spent time in the trenches of electrical engineering LTSpice is not exactly a household name No definitely not But if you're building any physical electronic device it's basically your gospel It's a high performance simulation tool Before you spend millions of dollars spinning up a physical prototype you use LTSpice to simulate exactly how electrical currents will flow through your circuit boards Precisely It stands for a simulation program with integrated circuit emphasis And it specifically excels at modeling switching regulators and power conversion circuits Which is key here Right that detail is critical Apple explicitly alleges the schematic was being used for power conversion development When we talk about

power conversion in consumer electronics we are talking about the absolute bottleneck of modern hardware design Because every device you own your phone your laptop draws power from a battery or a wall outlet But the internal components the microprocessors and the sensors require entirely different highly precise voltage levels And if you feed the wrong voltage to a sensitive chip you don't just get a software error No you literally fry the silicon Exactly And the byproduct of stepping that power up and down is heat Thermal management is the silent killer of hardware If you can't efficiently convert and manage power your device overheats or the battery drains

in 10 minutes Or the casing literally melts So finding the perfect mathematical balance to convert that power efficiently is incredibly tedious work It requires meticulous iteration Traditionally an electrical engineer will open a schematic in LTSpice run a simulation and just stare at the resulting waveforms Right They'll look at the voltage ripple analyze the thermal output and then manually tweak a single variable like adjusting a resistor value or modifying what we call a compensation parameter And a compensation parameter helps stabilize a feedback loop right So the power supply doesn't just fail when the system suddenly demands more energy Spot on The engineer tweaks that parameter runs

the simulation again and repeats this cycle hundreds or even thousands of times It's the ultimate trial and error And this is exactly where the allegations take that wild automated turn Apple isn't claiming that Liu sat at his desk for weeks manually tweaking those parameters They're alleging that he trained an AI agent to do it for him The claim is that an AI was instructed to run the LTSpice simulation visually inspect the results mathematically tune the compensation parameter and then run it again Which is just a staggering leap in workflow capability I mean we've seen AI write code we've seen it draft emails but LTSpice is

a complex desktop bound engineering environment It's designed strictly for human interaction Yeah To train an AI agent to operate that software analyze those highly specialized outputs and iteratively optimize the circuit design That is an incredibly advanced use of reinforcement learning It's like taking the stolen math homework and having the robot finish the rest of the equations Exactly If Apple's allegations hold true we're looking at the weaponization of artificial intelligence to not just utilize stolen IP but to accelerate it at machine speed The audacity is just off the charts But the filing doesn't stop at the technical workflow does it No it doesn't Apple is also

levying some very serious concealment allegations They claim that once Liu got wind of the internal investigation he went to a colleague at OpenAI and explicitly instructed them to destroy evidence And Apple claims that colleague agreed to do it That transforms this legal battle entirely How so Well it moves the dispute from a civil intellectual property disagreement where a company says hey you're using our patents into the realm of active obstruction In the eyes of a federal judge the alleged destruction of evidence often carries a legal presumption that the destroyed materials were in fact damning Furthermore Apple's filing alleges that multiple individuals within OpenAI were fully

aware of Liu's unauthorized access to Apple's third party cloud storage systems And to prove all of this Apple is zeroing in on the hardware They're demanding immediate access to a specific Mac mini along with a few other devices that are reportedly still in Liu's possession because this Mac mini was supposedly synchronized with the MacBook they already recovered And forensic synchronization in the Apple ecosystem is incredibly comprehensive Through systems like iCloud and local device handshakes a synchronized machine doesn't just copy files it mirrors state So it knows exactly what you're doing Exactly That Mac mini could contain a treasure trove of cached application states terminal logs

hidden version histories For investigators that machine could definitively map out the exact timeline When the file was opened which IP addresses the AI agent communicated with and how deeply this schematic was integrated into OpenAI's internal network So if you are OpenAI right now looking at a highly anticipated hardware roadmap this is a nightmare scenario because Apple isn't just asking for money No they're asking Judge Edward J Davila for an injunction that would outright bar OpenAI from using these trade secrets And an injunction is the nuclear option It doesn't just penalize past behavior it dictates future business operations This lawsuit prominently involves Tang Tan a former

senior Apple executive who left to join OpenAI's hardware efforts And we know through industry rumors that OpenAI is actively developing a novel screenless AI driven device Exactly And hardware development is intensely modular but power architecture is foundational Right you can't just surgically extract the power supply from a device that is already deep in the prototyping phase and just swap it out like a battery Right Everything else the thermal chassis the sensor placement it's all built around that specific power footprint Exactly If the court finds that Apple's proprietary power conversion logic is baked into OpenAI's upcoming hardware an injunction could force OpenAI to scrap the entire

design and start from scratch Oh man that could delay their entry into the consumer hardware market by years It could but you know we must heavily emphasize the procedural reality here These are unresolved allegations aggressively framed by Apple's legal team They are not independent findings of fact by a court Right and OpenAI is pushing back hard They are actively seeking the complete dismissal of the lawsuit calling the allegations meritless They stated on the record they do not possess Apple's trade secrets and they have absolutely no desire to acquire them Which sets the stage for a brutal evidentiary fight The immediate legal hurdle in front of

Judge Davila isn't whether the theft occurred it's over Apple's request for expedited discovery Right they want that Mac mini now Because normally discovery takes months or years Apple is arguing they need emergency access today because they believe the trade secrets are currently in active use and more importantly they fear the evidence is actively being destroyed We'll see how Judge Davila rules on that discovery request in October Next up staying with OpenAI the company is reportedly testing a radically different pricing model pay on completion Yeah this is a profound philosophical shift in how the software industry values its products For the last two decades the software

as a service or a SaaS industry has been built on a very simple premise You pay for access Right seat licenses or monthly subscriptions Or you know in the cloud era you pay for compute usage by the millisecond But in all of those models the vendor provides the tool and you the customer assume 100 of the risk of making that tool actually productive for your business But according to a new report from the information OpenAI is flipping that risk model on its head They are testing contracts where the enterprise customer only pays if the AI agent successfully completes a predefined task Exactly The classic example

being tested is a customer support ticket If a customer writes in and the AI agent fully resolves the issue without human intervention OpenAI gets paid for that resolution And if the AI gets confused or has to route the ticket to a human employee the customer pays nothing It moves the unit of sale from software access to verified work And we have to look at the broader enterprise environment to understand why this is happening right now Generative AI is astronomically expensive to operate compared to traditional cloud software Oh for sure The inference costs are huge Right Companies have been paying these massive usage based API bills

for a couple of years now Imagine you are the CFO of a Fortune 500 company You approved millions of dollars in AI pilot programs because it was the shiny new thing But now you're looking at your balance sheet and you are demanding hard ROI You're asking where's the revenue growth Where are the cost savings And if they can't point to definitive numbers those AI budgets are going to get slashed Outcome pricing is the perfect antidote to that CFO skepticism Because it transforms an abstract software expense into a measurable business outcome Exactly If a human agent costs you 8 to resolve a ticket and OpenAI offers

to resolve it for 2 but only if they succeed that is a proposition a CFO can instantly approve It's a no brainer And OpenAI is not making this move in a vacuum The competitive pressure here is immense Newer AI startups like Sierra and Finn have already built their entire go to market strategy around charging strictly for autonomous task completion And some companies are taking it to absolute extremes Look at a startup like Cognition They're literally offering corporate customers up to 10 million in credits If their AI software fails to deliver results that are equal to the purchase price That isn't just a money back guarantee

That is placing a massive financial wager on the competence of your own models And the legacy giants are following suit Salesforce just rolled out their AgentForce platform allowing enterprise clients to negotiate prices tied directly to revenue gains or lower service costs But you know while this sounds like a utopian alignment of incentives the operational reality introduces a massive friction point attribution How do you definitively prove that the AI caused the outcome In a binary scenario like resetting a forgotten password it's easy The user couldn't log in the AI reset the password they logged in task completed But what happens when Salesforce's AI is tasked with

negotiating contracts or driving sales revenue Right like let's use a real world marketing analogy Say your company spends 5 million on a massive national television and digital ad campaign A customer sees the ad they go to your website and they interact with an AI chatbot for two minutes just to find the checkout page And then they buy a 50 000 enterprise package Right so did the AI agent close that deal and earn a massive commission Or did your 5 million marketing team do all the heavy lifting Because if you're the head of marketing you are going to fight tooth and nail to say the AI

just acted as a glorified cash register That ambiguity is the Achilles heel of the pay on completion model Stripe recently issued a very stark warning about this exact dynamic Yeah they warned that seasonality marketing and product changes complicate attribution Exactly If your sales spike in November is it because your new AI sales agent is a genius or is it simply Black Friday demand What if your product team releases a highly anticipated new feature on the exact same day you deploy the AI You end up with multiple departments claiming credit for the exact same dollar of revenue This is where contract design becomes just as important

as the NOR network architecture Unless the vendor and the enterprise client agree in advance on ironclad rules of causation these deals are just inviting endless billing disputes Mopin AI has declined to comment publicly on these deals and the exact terms remain highly guarded But it's clear the industry is pivoting So the true product isn't the model it's the contract In research news a newly presented AI ECG system can flag heart risks in less than two seconds And the scale and the rigor of the data backing this development are what demand our attention here The findings were just presented in Munich at the annual Congress of

the European Society of Cardiology Okay The analysis was spearheaded by researchers at Imperial College London They trained this AI system on millions of routine electrocardiograms or ECGs and then tested it in a massive United States trial involving 67 000 patients Let's break down the mechanics of this because if you've ever been to a hospital for chest pain you know the ECG They stick the little adhesive nodes to your chest and limbs and it prints out a graph with those rhythmic squiggly lines It's just recording the electrical activity of your heart Exactly It's a phenomenal tool for detecting electrical abnormalities like arrhythmias or an active heart

attack But historically an ECG is essentially blind to the physical structural integrity of the heart Like a valve is failing Right To diagnose structural issues like heart failure or valve disease cardiologists have always relied on an echocardiogram Which is an ultrasound You need specialized equipment a trained sonographer and a cardiologist to interpret the video If you're in a healthcare system right now the waiting list for an outpatient echocardiogram can easily stretch for several months And this logistical bottleneck is exactly what this new AI system is designed to attack The researchers demonstrated that their AI can look at the simple electrical trace of a standard ECG

and detect subtle microscopic patterns that correlate with structural heart failure and valve disease Patterns that human eyes can't even see Exactly In that 67 000 patient trial the AI successfully identified up to 81 of the patients who actually had heart failure and up to 90 of the patients suffering from heart valve disease just from reading the electrical trace That is a staggering capability It's effectively using a cheap electrical test to predict the results of an expensive structural test in under two seconds So the proposed role here is a massive triage layer You run this AI on routine hospital ECGs to surface unsuspected diseases But we

have to address the immediate logistical pushback here Right because if you flag everyone in two seconds doesn't the waiting room for the actual ultrasound just get more crowded You're just jamming more people into a pipeline that's already choked It's the most vital question in medical AI deployment You are absolutely right that you aren't increasing the supply of echocardiograms but what you are doing is drastically optimizing the prioritization of the demand Under the current system a months long wait list is often managed on a first come first served basis By applying this AI triage layer you ensure that the patient who is scheduled to wait three

months is the patient who can actually afford to wait three months without dying And the patient whose AI readings suggest their valve is failing gets bumped to the front of the line for a scan tomorrow morning It changes the waiting list from chronological strictly risk based But we have to underline the absolute most important caveat of this research These results are strictly classified as case finding Right meaning it's a probabilistic warning system It is absolutely not a definitive clinical diagnosis The AI cannot diagnose or rule out either condition and echocardiogram is still strictly required for confirmation Exactly And according to Ahmed Elmadani the clinical research

fellow who led this analysis the next major hurdle is hardware integration His team wants to design handheld AI powered ECG readers for clinicians A great triage tool but not a standalone doctor yet Turning to open source tools OpenClaw has released version 2 0 adding shared agent sessions but explicitly skipping tenant isolation For anyone tracking the frantic evolution of open source AI agents the release cadence of OpenClaw has been astonishing They had 106 distinct releases in the 230 days prior That's a blistering speed And then everything just stopped There was a highly notable seven week pause before they finally dropped this massive 2 0 update In

the open source world a seven week silence usually means they're rewriting the very foundation And looking at the changelog they definitely did They introduced a rebuilt browser control UI they migrated session data into Squalate and added shared cloud sessions Moving to Squalate is a highly strategic architectural decision By moving session state to Squalate OpenClaw 2 0 makes the entire agent environment incredibly portable You can literally drag and drop your agent's memory from your laptop to a cloud server in one second And they paired that with a completely model neutral setup process It automatically tests and reuses existing sign ins for OpenAI's codecs ChatGPT Clawed or

even local models like Allama and LM Studio Which is seamless You can route your agent's logic through a massive hosted frontier model for complex tasks and then switch to a small local model for highly sensitive source code Plus the performance bumps are huge Simulated startup dropped from 1 6 seconds down to just 575 milliseconds And JavaScript requests fell from 140 to 45 making that new UI super snappy But the headline feature is the introduction of shared cloud sessions It radically lowers the friction for collaboration A second developer on your team can literally join the live session or even take over the controls without losing any

of the historical context But and this is the huge caveat the shared sessions are not tenant isolation Right some boxing remains off by default Let me translate this into real world terms It's like having the key to a house Giving your best friend a copy of the key so they can come over and share the kitchen is fantastic That's what OpenClaw 2 0 does for a trusted team But you absolutely would not give that exact same key to a random stranger you're renting a room to Especially if there are no locks on any of the interior doors That is the perfect analogy Tenant isolation is

the equivalent of those interior locks OpenClaw 2 0 is designed explicitly for one highly trusted operator or a tightly knit internal team It is absolutely not a safe foundation for a multi tenant SaaS platform where hostile customers share resources The operator retains all responsibility for the trust boundary so you have to manage the blast radius yourself Great for colleagues bad for a SaaS platform Let's move into our quick reads First up OpenAI is reportedly buying tens of thousands of Mac minis and Mac studios for computer use agent training And this is specifically for reinforcement learning training agents to operate computers across multi step tasks Navigating

GUIs clicking buttons that sort of thing What's interesting is the contrast Anthropic is doing similar research but they rent Mac minis via AWS OpenAI is buying the hardware directly Why Macs though when NVIDIA GPUs are the kings of AI It comes down to Apple's unified memory architecture The CPU GPU and neural engine all share the same massive pool of high speed memory For the specific workload of training an AI agent to read a screen hold that visual context and calculate an action that unified memory is a huge advantage over traditional PCIe bus transfers But the caveat here is the supply chain High end Macs have

been sold out for months due to a memory chip shortage Plus as you noted before we started there's no public data comparing the cost to performance ratio of these Mac clusters against NVIDIA GPUs for this specific workload Right it's a very niche hardware grab Quick read two A real world test of Unitry's new 4 007 taller Go 2 Pro robot dog shows that legged robots are getting cheaper but they aren't everyday tools yet Yeah a buyer documented their experience The robot successfully navigated a two mile downhill trip But on the uphill return journey it completely collapsed It had 5 battery left and an internal motor

temperature of 84 degrees Celsius It cooked itself walking up a hill And this comes down to the gear reduction right Exactly Unitry pushes the price down with aggressive modularity using 12 interchangeable motors and a very low 6 3 to one gear reduction ratio Boston Dynamic Spot uses up to a 51 to one ratio So the G2's motors are just working absolute overtime to fight gravity Yes which instantly translates into severe heat But the financial context is what matters Following its recent IPO Unitry's valuation hit 34 billion They're pushing even cheaper humanoid robots now like the G1 starting at 13 500 Even if the robot gives

up on a hill Plus without arms or hands wheels or drones are often just better anyway Finally our third quick read A proposed California class action by Karl Kahn claims Anthropic's 200 CloudMax plans deliver significantly less weekly usage than advertised The dispute revolves around how the limits are communicated Anthropic advertises 5X and 20X usage per five hour session But Kahn's complaint alleges that when you look at the total weekly estimates like 240 to 480 hours of access to Sonnet 4 that only equates to roughly 6X the weekly capacity of the standard pro tier not 20X And tools like Cloud Code share this exact same allowance

pool so automated scripts can drain it fast Right but we must be incredibly firm here These are the plaintiff's calculations based on company estimates These are not independent measurements and it is absolutely not a court finding No class has been certified yet That covers the news Number one Apple's new evidence against OpenAI escalates the trade secret fight highlighting the bizarre alleged use of an AI agent to tune a stolen schematic Number two software pricing may be shifting from access to outcomes as OpenAI reportedly tests charging only for successfully completed tasks Number three hardware capacity for AI is branching out with OpenAI buying up Macs for

agent training while cheap robotic dogs still struggle with basic real world endurance And the single development worth watching tomorrow Tomorrow keep an eye on how enterprise software buyers react to these pay on completion AI deals The fight over attribution rules is just beginning You can find every story we covered and more at superpowerdaily com Thanks for listening to the Deep Dive and we'll see you tomorrow