ChatGPT adds interactive charts and calculators

Welcome to The Signal from Superpower Daily with Maya and Theo Right now ChatGPT is moving beyond simple text answers to become a fully interactive workspace And that fundamentally changes what you can expect from an AI It really does So today we are jumping straight into this deep dive We are looking at a massive shift in how we interact with these models Yeah The conversational engine is becoming a dynamic tool It is for everyday planning It is for learning And the controls are being built directly into the answers themselves Which is wild to think about Today is really about the growing pains of trusting and interacting

with AI We are looking at everything from what we see on our screens to the medical data hidden in our hospitals So let's unpack exactly what is happening with this new OpenAI rollout Let's do it They are calling this Intelligent UI Right And they are embedding editable graphs right into their responses Yeah They are adding interactive charts I mean they are even putting task specific calculators directly into the chat window We actually saw this demonstrated recently It was by product manager Arush Selvan Oh yeah I saw that The examples were genuinely fascinating Selvan asked the model to explain something pretty complex He asked how an

airplane wing creates lift Which is a heavy physics question Exactly Normally you would get a wall of text It would explain fluid dynamics It would mention Bernoulli's principle Which most people just skim over I know I do Or they ask the AI to summarize it again in simpler terms Right Because reading a textbook chapter in a chat window is exhausting It really is But instead of text the system generated an interactive diagram It was a visual of an airplane wing Wow Yeah You could visually see the mechanics You could actually see the airflow over the curve of the wing That is incredible They showed other

stuff too They did recipe visuals Oh nice They did a breakdown of bicycle mechanics They even generated a map for a multi day hike Wait like an actual map you can interact with Yes And they built a personal savings calculator right there in the window So this shifts the entire paradigm I mean of interacting with a large language model It completely changes it Because you are no longer just reading an answer You are manipulating a workspace Exactly If the AI gives you a graph you can tweak the parameters right there You do not have to write a whole new prompt you know just to change

a single variable And that is the core technical leap here Break that down for me Well think about how a language model usually works It just predicts the next word in a sequence Right It generates text tokens Exactly But now it is generating code And that code renders an interactive user interface module in real time That is a huge jump It is It turns the AI from an oracle that hands you a static answer into a collaborative tool You are working with the data now You are not just consuming it I want to dig into how that actually changes the user experience though OK Because

it sounds like it could be a bit overwhelming Oh for sure Is this like asking someone for the time and instead they hand you a schematic for a Swiss watch That is a great analogy Right Because if I just want a quick text answer does an interactive interface ruin that Does it turn my quick chat into a cumbersome dashboard That is a very real concern Visual clutter is a notorious problem in user interface design It really is If you bombard a user with sliders and charts for a simple question you ruin the user experience You increase their cognitive load You do not reduce it Which

defeats the entire purpose of a quick AI assistant Exactly But OpenAI seems to recognize that They know not every query requires an interactive dashboard Oh good So they built in a control for this exact issue Users who prefer a text heavy interface can dial back the visuals OK That makes sense It basically operates on a sliding scale of complexity It is based entirely on your preference I am very glad to hear that Because sometimes you just need a simple sentence Right Just tell me the answer Exactly But I do see the value of a task specific calculator especially for complex planning Oh absolutely Trying to

work out compound interest purely through text prompts gets incredibly tedious you know going back and forth Yeah Think about that hiking map example Right When you are balancing multiple variables human cognition struggles You have terrain You have daily mileage You have elevation changes That is a lot to keep in your head It is too much We benefit heavily from spatial and visual organization So the calculator or the map format offloads that cognitive burden I see It moves the effort from your working memory into the interface You can adjust your daily mileage on the visual hiking map and then you instantly see how it changes your

campsite locations That is so cool That makes the AI feel much more like an operating system It really is a fundamental shift in computing So we should talk about the rollout schedule for this because it is highly staggered It is very fast but staggered So Pro Plus Business and Enterprise accounts started getting this on October 7 Yep But Free and Go users follow the very next day on October 8 That rapid succession is super interesting They are putting the tool in the hands of millions almost immediately But we have to look at the reporting around this rollout too There is a massive caveat floating around

the tech press right now Ah yes The rumor mill TechCrunch reported that this launch coincides with the GPT 6 Yeah that reporting caused a lot of stir online But there is a crucial detail omitted from that TechCrunch piece What is that The actual year Oh wow Yeah They stated an October date for GPT 6 But without specifying the year we cannot treat that as a confirmed immediate release That could be October of next year Exactly We cannot assume it means the underlying model is updating today The interface update however is confirmed That is happening right now That is a very important distinction to make We

are talking about a UI wrapper update here not necessarily a completely new foundational model Correct So what should we be watching next with this story You should keep a very close eye on that October 8 rollout the one for the free and go tiers OK Why that one specifically We need to see how millions of general users respond to this interactive format Paid tier power users usually love dense features They want all the dials and knobs That is true Right They are paying for it Right But general users might hate them It will be very telling to see if everyday people actually use the visual

controls Or if they just get annoyed by them Exactly Will they dive into the settings and turn the interface back into a simple text chat The adoption rate will really tell us if people actually want an AI workspace It will reveal a lot Are we ready for a new computing paradigm Or do we just want a smarter search box Let's close the book on OpenAI's interface shift Next up moving from generating new visuals to verifying where they actually came from Google has officially opened its AI watermark checker globally This is a major democratization of watermark detection It sounds huge It is Google just launched a

dedicated site It is at SynthID com And it is available globally in English Okay This brings the detection capability entirely out of Gemini and Google Search It turns it into a standalone utility for the public And the scale of this deployment is just staggering SynthID is already embedded in over 180 billion images and videos It is a massive number They have also watermarked 240 000 years of audio content That footprint is wild It shows how aggressively Google has been watermarking media in the background before we even really noticed But the site checks media for invisible pixel or waveform signals I want to stop here and

ask how that actually works Sure Because we are not talking about like a Getty Images logo stamped across a photo right No not at all Visible watermarks are basically useless for security Because anyone can just crop them out Or blur them SynthID uses steganography Steganography Okay what is that It alters the fundamental noise pattern of the file itself In an image it changes the actual pixel values But it does it in ways the human eye cannot detect Oh wow And in audio it subtly shifts the frequencies just a tiny bit So if I take an AI generated image and I crop it is the watermark

still there Yes The pattern is distributed throughout the entire file Even if you crop it or compress it or apply a crazy color filter The watermark survives It survives The underlying mathematical signature remains perfectly intact The detector algorithm just scans the file It looks for that specific mathematical pattern And if it finds it it flags it That is brilliant It is very clever engineering Right now this standalone site supports watermarks from Google OpenAI NVIDIA and Kakao Yep those are the big ones currently supported Apple support is planned for later But there is a catch with this tool There is always a catch You have to

sign in to use the tool And there is a very strict limit on usage Users get approximately 10 checks per day across all formats That is an incredibly tight bottleneck It really is And just to clarify that is not 10 for images and another 10 for audio It is 10 total checks per day per account That feels way too low for a global utility I mean think about it If I am a journalist trying to verify a batch of photos from a news event Right I will burn through 10 checks in five minutes Why throttle it so tightly The 10 check limit exists explicitly to

deter bad actors How does limiting checks stop bad actors Well if you give someone unlimited access to a detection tool they can use it to reverse engineer the watermark Wait how would they do that They would run automated scripts They take an AI generated image They subtly alter a few pixels Then they run it through the detector OK If it still flags as AI they alter it again And they just repeat this thousands of times It is a process called gradient descent Oh I see Yeah they do this until the detector finally fails to trigger So they are basically using the detector as a training

mechanism To learn how to wash the watermark out completely It becomes an adversarial training loop And once they find the exact combination of alterations that break the detector they are golden They can build software to strip the watermark entirely from any file Wow OK so the 10 check limit breaks that automated loop Right You cannot run thousands of automated tests if your account gets locked after 10 tries That makes total sense from a security standpoint But the public tool also has another limitation especially compared to what professional media organizations get Yes the internal tools are much more robust Right The internal tool highlights the specific

manipulated regions of an image If someone took a real photo and used AI to add a fake building in the background the internal tool draws a box around the fake building It shows you exactly where the manipulation happened But this public site will not do that It just gives you a basic text message saying AI was detected Correct It is a simple binary answer It tells you if a supported signal is present or not And that actually brings us to the most critical caveat of this entire system What is that A negative result is absolutely not an all clear OK let me make sure I

understand this It is like scanning a crowd for people wearing a specific brand of red hat Right If you do not see any red hats it does not mean there are no spies in the crowd It just means they are not wearing that exact hat That is the perfect way to visualize it The tool cannot definitively prove a file is human made Because it is only looking for specific brands of red hats Exactly The media might be generated by an AI system from a smaller company one that is not supported by SynthID Right Or the creator might have used an open source AI model and

that model applies no watermark at all So a negative result simply means the tool did not find what it was specifically programmed to look for It absolutely does not mean the photo is real It is a helpful tool but it is not a universal shield against deepfakes So what should listeners watch next with this SynthID rollout Watch for the upcoming Apple integration OK We really need to see if other major providers adopt the SynthID standard Why is that the key metric Because the utility of this tool scales entirely with industry adoption Until every major image and audio generator embeds these specific signals the tool is

limited The absence of a watermark will never guarantee authenticity until everyone is on board We will definitely keep tracking that industry adoption In other news we are shifting from verifying public media to securing highly sensitive private health data This next story is intense It is Epic has paused most of its new product development to address critical security flaws Epic is a massive medical records giant just to give some context Right CEO Judy Faulkner noted this pause would likely last six weeks A six week pause in development is huge for a company that size It is The security flaws were found in some MyChart customer configurations

and they were discovered by an AI system It was Anthropic's Claude Mythos This story is a perfect illustration of the double edged sword of a shared software platform Let's talk about the scale first Epic holds over 320 million patient records They control nearly 44 of the acute care electronic health record market That footprint is staggering It really is It is a massive centralization of incredibly sensitive data Which means a centralized vendor can use advanced AI to find vulnerabilities incredibly fast right Yes that is the upside Anthropic's Claude Mythos was able to analyze the code It spotted structural weaknesses that human auditors completely missed But there

is a massive downside Standard configurations mean a single discovered vulnerability has a massive blast radius Exactly This is the danger of a software monoculture Explain that term Well think about a housing development If you find a structural flaw in a custom built house you only have to fix one house Right But if you find a flaw in the blueprint of a mass produced housing development you suddenly have a neighborhood wide crisis Everyone's roof might cave in That makes total sense I want to talk about how an AI actually finds these flaws though It is not just reading the code like a human right No it's

much more aggressive than that It uses techniques like static code analysis and automated fuzzing Yeah It basically throws thousands of unexpected inputs at the software just to see how it breaks Wow It maps out the logic flow of the application to find loopholes Cloud Mythos acts as an automated red team It thinks like an attacker It maps out potential pathways into the system The actual technical details of the bugs remain undisclosed obviously Yeah they are keeping that very quiet to prevent exploitation But Epic security chief Sterling Martin shared a very concerning detail The flaw might allow access to records without triggering an intrusion log That

lack of an intrusion log is terrifying That is what elevated this from a routine patch to a full product development pause Why is the log so important In medical data auditing is everything Because of HIPAA Exactly HIPAA compliance requires strict tracking of who looks at what record If an attacker gains access and trips an alarm security teams can respond They can shut it down Right They can lock down the system They can notify the affected patients But if an attacker gains access invisibly you have no way to audit what they viewed You do not even know you have been breached Exactly This is like finding

out the security cameras at the bank were turned off for a month That is a great way to put it We do not know if anyone actually walked into the vault and altered the ledger We might count the money and find it is all there Right But the fact that the guard wouldn't even have a log of the visit is what makes everyone freeze That is exactly it And crucially the AI did not establish whether an attacker could actually change those patient records undetected Wait really Yeah It only identified the unlogged access route It showed the door was unlocked and the camera was blind But

it didn't prove anyone went in and changed files Right But in health care unlogged access is a severe compliance and privacy violation on its own We also do not have a firmly confirmed end date for this pause Faulkner estimated six weeks back in September The operational strain on hospitals is very real right now Yeah You have hospitals trying to improve security but they cannot cut clinicians off from resources critical to active patient care That is the tightrope health systems walk every single day You cannot just reboot a hospital You cannot simply take an acute care hospital offline to patch a system Patients are in surgery

The emergency room is full Doctors need constant access Uninterrupted access They need patient histories medication lists and lab results right that second So the logistics of deploying these critical security fixes across 44 percent of the market while maintaining uptime is just staggering It requires immense coordination between EPIC engineers and local hospital IT teams What should listeners watch next regarding this EPIC pause Watch exactly that operational balancing act Look for reports from hospital networks on how they are managing to implement these vital security patches without disrupting patient care It sounds like a massive test for the industry It is a profound test of how resilient centralized

health care infrastructure truly is especially when faced with zero day vulnerability management We will definitely see how they manage that six week timeline Meanwhile let's look at the financial sector Trust issues are creating a massive bottleneck in corporate workflows right now Yeah this data is wild DataRails just released a fascinating survey of 270 U S finance executives And these are at large organizations right The survey targeted organizations with over 100 million in revenue Okay so big players Very big And the headline finding is crazy These financial leaders are spending 26 percent of their week checking or correcting AI outputs That is more than a quarter

of their working week devoted purely to verifying machine generated work It is a massive time sink Furthermore 96 percent of these executives spend at least a tenth of their time fact checking The trust level across the industry is incredibly low It is practically non existent at the top Only 5 percent trust AI for board ready reports without human review Yet despite this massive time sink 53 percent still plan to add more AI licenses next year I know That feels like a total paradox Why buy more of a tool that is eating up a quarter of your week Because the FOMO is real Fear of missing

out Fear of missing out drives a lot of enterprise software purchasing That makes sense But this survey exposes a critical workflow and governance gap in enterprise AI Organizations are buying the software because they want the promised efficiency But they lack the operational readiness to actually trust the output Exactly And the core issue isn't even the AI itself What is it It is the underlying data architecture of these companies Oh right The AI is just the top layer The foundation is completely broken The numbers on that foundation from the survey are staggering Only 4 percent of respondents have a single source of truth for their finance

and operational data Think about that Only 4 percent Meanwhile 23 percent rely on completely disconnected systems and manual reconciliation Let's explain what disconnected systems look like in a real company Okay lay it out The marketing team uses Salesforce to track leads The finance team uses NetSuite to track revenue And the operations team has this custom SQL database built 10 years ago to track inventory And none of these systems talk to each other seamlessly Exactly They are siloed So when you plug an AI into that mess what happens Total chaos Pretty much The AI uses a process called Retrieval Augmented Generation or RAG Right It searches

through all those internal databases to find the answer to your prompt If you ask it for the quarterly revenue projection it pulls conflicting numbers from Salesforce from NetSuite and from the old SQL database And it does not know which one is the authoritative answer because the company doesn't even know Right It blends them together Or it picks the most recent document even if it is just a draft Oh that is dangerous It is Then it hallucinates a very confident completely incorrect conclusion Which explains another stat from the survey 65 of respondents noted that their biggest frustration is AI giving confident answers based on wrong data

Yep It sounds like they have hired a hyper confident intern who speed reads the wrong filing cabinets That is exactly what it's like If you are spending a quarter of your week grading the intern's homework are you actually saving any time Or are you just shifting your job title from accountant to AI babysitter You are absolutely just babysitting the AI at that point An AI model is essentially a reasoning engine OK If you feed that engine contradictory spreadsheets the reasoning engine will fail Garbage in garbage out But with AI it is garbage in highly articulate and persuasive garbage out Exactly That's why it takes 26

of their week to untangle it They have to trace the AI's logic back to the original flawed data source We do need to highlight the limitations of this data though Yes definitely This was an online survey conducted in July It is entirely self reported by these executives It represents a specific cohort of large companies It is not an industry wide objective measurement of actual hours logged via time tracking software That is a very fair caveat We also must be careful with causality here What do you mean The correlation between disconnected systems and AI hallucinations is a strong association in this survey But it is not

definitive proof of cause Meaning other factors play a role too Yes The specific AI models used matter A cheaper smaller model will hallucinate way more than a frontier model That is true Also the prompt engineering skills of the finance staff play a huge role If they write vague prompts they will get useless answers regardless of how clean the underlying data is Still the hurdle for enterprise adoption is very clear 75 of respondents cited auditability as a major barrier What should listeners watch next in this space Watch whether organizations can fix their underlying data consolidation Only 7 of these leaders say they are fully ready for

AI across their workflows So their budgets are vastly outpacing their actual readiness Yes They need to clean their data lakes before they buy more AI fishing pools I love that Until companies build a single source of truth they will continue paying for expensive AI licenses only to spend their Fridays double checking the machine's math We are shifting to our quick reads now Chick fil A is officially keeping humans at the drive thru Really Yes CEO Andrew Cathy stated the chain will not use AI for drive thru voice ordering They are actively opting for a human plus approach Cathy wants to protect their high ranking hospitality

They are famous for their customer service This decision is a very sharp contrast to the rest of the fast food industry McDonald's is currently testing the Archie voice system And Wendy's is already rolling out their fresh AI assistant It is a bold strategic divergence Chick fil A is explicitly drawing a boundary around customer facing automation They might use AI in the kitchen or for supply chain logistics though Oh certainly But they believe the human interaction at the window is a core part of their brand value Next up OpenAI and Chip Ganassi Racing just released episode two of a new R D documentary This is pretty

cool The film showcases AI's role in Alex Pillow's IndyCar championship season And IndyCar produces information from hundreds of sensors It tracks tire temperature downforce fuel mixture and suspension telemetry The volume of data is just overwhelming for a human engineer to process in real time during a race weekend So the AI sifts through that massive volume of telemetry It finds useful signals for car setups Right It is accelerating the engineer's judgment It highlights anomalies and suggests setup changes But it is not replacing the human engineers No not at all However there is a major caveat in this documentary What is that The team has not actually

quantified the time saved They also have not provided the AI's exact statistical contribution to those specific race wins So it is a fascinating use case for data sifting but the exact return on investment remains unstated Exactly Finally Michael Smith gets 18 months in prison for AI music streaming fraud This is a landmark ruling The 54 year old was sentenced to 18 months He has been forced to forfeit over 8 million And he was given two years of supervised release Since 2017 Smith used bots to play hundreds of thousands of AI generated tracks billions of times He targeted platforms like Spotify Apple Music and Amazon Music

He deliberately spread the fake streams across a vast catalog Right He generated thousands of generic tracks using AI music generators Then he programmed his bot networks to listen to them But he kept the play counts on each individual track low To evade fraud detection algorithms Exactly The scale of the bot network is incredible In April 2023 alone his bots generated 80 9 million plays on YouTube music family plans That is insane For comparison Taylor Swift's entire catalog had 9 3 million plays in that same category during that same month It's mind blowing This is a highly orchestrated draining of shared royalty pools He stole 8

million from legitimate artists by artificially inflating his slice of the streaming pie It shows how AI generation combined with automated bot networks can just break traditional compensation models We are moving to three takeaways from today Okay First the definition of an AI chat interface is expanding Open AI is embedding interactive editable tools directly into answers This moves the technology far beyond static text into a dynamic workspace Second trust and verification are becoming massive bottlenecks across all industries We see this with Google strictly capping watermark checks to prevent abuse We also see finance executives losing a quarter of their week double checking AI outputs because their

underlying data is a mess Third the human in the loop is proving incredibly resilient Whether it is Chick fil A protecting customer hospitality at the drive thru or Ganassi engineers using AI purely as a data sifting assist human judgment remains the premium asset in an automated world Here is one development to watch tomorrow Pay close attention to how general users react to the OpenAI Intelligent UI rollout when it hits the free and go tiers on October 8th We need to see if the everyday user embraces the interactive dashboard or if they find the visual clutter overwhelming Head over to superpowerdaily com for more on all

these stories Thank you for listening We'll see you tomorrow