Tech Brew Ride Home - Fri. 10/28 – GPT-4.5

Episode Date: February 28, 2025

OpenAI has launched GPT-4.5, but they want you to know, this is not a frontier model, it’s a stepping stone. Skype enters the Deadpool. Memecoins are ok by the SEC. Most criminality in crypto is now done with stablecoins. And in the Weekend Longreads Suggestions, how the MTA is using Pixel phones to do subway maintenance. Links: OpenAI announces GPT-4.5, warns it’s not a frontier AI model (The Verge) OpenAI CEO Sam Altman says the company is ‘out of GPUs’ (TechCrunch) Microsoft hangs up on Skype: service to shut down May 5, 2025 (TechCrunch) Microsoft is killing Skype after 14 years of neglect (Windows Central) Meta plans to release standalone Meta AI app in effort to compete with OpenAI’s ChatGPT (CNBC) SEC says most meme coins are not securities (CNBC) Illicit Crypto Volume in 2024 Hit a Record $40B in 2024 (CoinDesk) Meta is firing about 20 employees for leaking information (The Verge) Weekend Longreads Suggestrions: The New York City Subway Is Using Google Pixels to Listen for Track Defects (Wired) Learn more about your ad choices. Visit megaphone.fm/adchoices

On April 4th, 2023, around 2 in the morning, a man was found stabbed multiple times on a sidewalk in downtown San Francisco. Hey, who did this to you? What happened next turned the story into a political firestorm. Reports have identified the victim as Bob Lee, the founder of Cash App. From Bloomberg Podcasts, this is Foundering, the Killing of Bob Lee, beginning April 16. Welcome to the Tech meme right home for Friday, February 28th, 2025. I'm Brian McCullough today. OpenAI has launched GPT4.5, but they want you to know this is not a frontier model. It's a stepping stone. Skype enters the Deadpool. Meme coins are okay by the SEC. Most criminality in crypto is now done with stable coins. And in the weekend long-reed suggestions, how the MTA is using pixel phones to do subway maintenance. Here's what you miss today in the world of tech. OpenAI has launched GPT 4.5, what it calls its most knowledgeable model. into research preview, initially warning it's not a frontier model and may perform below 01 or 03 mini. Quoting the verge, GPT4.5 will have better writing capabilities, improved world knowledge,

and what OpenAI calls a refined personality over previous models. Open AI says interacting with GPT4.5 will feel more natural, adding that the model is better at recognizing patterns and drawing connections, making it ideal for writing, programming, and solving practical problems. However, OpenAI notes it won't introduce enough new capabilities to be considered a frontier model. GPD4.5 is not a frontier model, but it is OpenAI's largest LLM improving on GPT4's computational efficiency by more than 10x. OpenAI said in a document that leaked ahead of its announcement, it does not introduce seven net new frontier capabilities compared to previous reasoning releases and its performances below that of 01, 03 Mini, and deep research on most preparedness evaluations. It was previously reported that OpenAI was using its 01 reasoning model codename strawberry to train GPT4.5 with synthetic data. OpenAI says it has trained GPT4.5 using new supervision techniques combined with traditional methods like supervised fine tuning and reinforcement learning from human feedback, similar to those used for GPT4O, end quote.

Sam Altman further said OpenAI was forced to stagger GPT4.5's rollout because it is out of GPUs, quoting TechCrunch. We've been growing a lot and are out of GPUs. as Altman wrote, we will add tens of thousands of GPUs next week and roll it out to the plus tier then. This isn't how we want to operate, but it's hard to perfectly predict growth surges that lead to GPU shortages. Perhaps in part due to its enormous size, GPD 4.5 is wildly expensive. OpenAI is charging $75 per million tokens fed into the model and $150 per million tokens generated by the model. That's 30x the input cost and 15x the output cost of OpenAI's Workhorse GPT40 model, end quote. Initial looks into GPD 4.5's claimed performance, including on coding benchmarks,

say it matches or outperforms GPT40, but falls short of OpenAI's deep research. Quoting TechCrunch, OpenAI plans to eventually combine its GPT series of models with its O reasoning series, beginning with GPT5 later this year. GPT 4.5, which reportedly was incredibly expensive to train, delayed several times, and failed to meet internal expectations, may not take the AI benchmark crown on its own, but OpenAI likely sees it as a stepping stone towards something far more powerful, end quote. So again, GPT 4.5, what the internet was calling Orion, is now available via the $200 per month chat GPT Pro subscription. Quoting Wired, chat GPT 4.5 is in stark contrast to other recent AI innovations like Deepseeks R1 that

attempted to match the performance of a frontier model with as few resources as possible. Open AI still sees a strong path forward through scaling its models, according to researchers who worked on GPT4.5, this kind of maximalist's mindset to model development has captured more of the nuances of human emotions and interactions. They see the model size as also potentially helping this iteration hallucinate less frequently than past releases. If you know more things, you don't need to make things up, said Mia Glaze, who leads OpenAIs alignment team and human data team.

Exactly how big or compute-intensive GPT4.5 is remains unclear. OpenAI declined to share specific numbers, end quote. Online discourse around this can be summed up by at J.M. underscore Alexia, quote, GPT 4.5 is, as I expected, just more pre-training data. Pre-training scaling for text is mostly saturated. Reasoning, O-1, Deepseek made us think that progress was going up exponentially, but it's mostly better at tuned benchmarks. Progress continues, but it's not exponential, end quote. Skype is dead. Microsoft says it will shut down Skype on May 5th and direct users to migrate all their contacts and chat data to teams. As recently as 2023, Skype had more than 36 million daily

active users, quoting TechCrunch. After kick-starting the market for making calls over the internet, 23 years ago, Skype is closing down. Microsoft, which acquired the messaging and calling app 14 years ago, said it will be retiring it from active duty on May 5th to double down on teams. Skype users have 10 weeks to decide what they want to do with their account. It's not clear how many people are impacted. The most recent numbers that Microsoft shared were in 2023 when it said it had more than 36 million users, a long way from Skype's peak of 300 million users. We know this is a big deal for our Skype users, and we're very grateful for their support of Skype and all the learnings that have factored into Teams over the last seven years. Jeff Tepper, president of Microsoft

365 collaborative apps and platforms, told TechCrunch in an interview this week. At this point, putting all our focus behind teams will let us give a simpler message and drive faster innovation, end quote. Between now and May 5th, users will have the option to migrate all their contacts and chat data over to Microsoft's Teams platform. Alternatively, users can download their Skype data using the app's built-in export tool, end quote. And quoting Windows Central. Skype in its current form is a sorry excuse for a messaging platform. It's a web-based app on Windows, which is very slow and does a terrible job at syncing

conversations and notifications across devices. The app also doesn't even support Windows on Arm, which is something Microsoft has been pushing to adopt over the last half a decade. In the last few years, Microsoft has tried to revive Skype with experiments such as Skype clips, which let users send short video messages to each other without calling them. In 2024, it also gained access to code. pilot, self- messaging, and a revamped mobile experience designed to compete with rival platforms. Nothing seems to have worked, as the service still hasn't seemed to have gained much traction.

The last time Skype received a notable update was in May 2024. Microsoft recently shuttered its Skype number service, which let you buy a real phone number that could be used to call other phone numbers, not on the Skype service, end quote. The sources say meta plans to launch a standalone meta-AI app in Q2 and test a paid subscription service for meta AI as old Zuck tries to make meta the leader in AI, quoting CNBC. Unlike competing generative AI tools such as chat GPT and perplexity, meta AI is currently only available to users via a website and the company's apps such as Facebook and WhatsApp.

Although meta's vast user base across its family of apps can access meta AI, users could potentially interact more deeply with the digital assistant if it were available as a standalone app, the people said. In January, Zuckerberg publicly agreed with a Threads user who said meta should create a standalone mobile app for its digital assistant. The Threads user wrote that a separate meta AI app could help the company unify the digital assistant across smartphones and different hardware platforms, such as the RayBan meta smart glasses, help users organize their conversational histories with the digital assistant and allow for

deeper personalization and customization. Zuckerberg responded to the thread user with a red 100 emoji, often used in online chatter to convey wholehearted agreement. Meta also plans to test a paid subscription service for meta AI similar to the way OpenAI and Microsoft charge users monthly fees to access more powerful versions of their respective chatGPT and co-pilot chatbots the people said, end quote. The SEC says that most meme coins are not securities, are instead analogous to collectibles, and meme coin purchasers and holders are not protected by federal securities laws,

quoting CNBC. It is the division's view that transactions in the types of meme coins described in the statement do not involve the offer and sale of securities under the federal securities laws. The statement says, persons who participate in the offer and sale of meme coins do not need to register their transactions with the commission. Accordingly, neither meme coin purchasers nor holders are protected by federal securities laws, end quote. It also said a meme coin does not constitute any of the common financial instruments specifically enumerated in the definition of security, because among other things, it does not guarantee a yield or convey rights to future income, profits, or assets of a business. In other words, a meme coin is not itself a security, end quote. The clarification comes after the latest rapid rise of such cryptocurrencies following the election of President Donald Trump, as well as their crash in recent weeks. It's also another notch in the belt of the new administration, which has promised to create clearer and perhaps more favorable regulatory conditions for the crypto industry and to do so swiftly. The clarity could pave the way for exchange operators to list more meme coins without the risk of regulatory enforcement. In January, at the height of the Trump-fueled meme main

Coinbase CEO Brian Armstrong said that given there are about one million tokens a week being created now and growing, evaluating each one by one is no longer feasible in a post on X. And regulators need to understand that applying for approval for each one is totally infeasible at this point, he said. Meme coins, of which there are thousands, sit at the furthest end of the risk spectrum. They're three to four times more actively traded than Bitcoin and Ether, adjusting for market cap, which makes them lucrative offerings for newcomers to the market who feel they may have missed the boat on Bitcoin. Historically, they've been a gauge of retail interest and risk appetite in crypto, though most market participants warn strongly against them. Despite their purely speculative

nature and lack of intrinsic value, they're widely viewed as a significant sector of the crypto market and an important part of internet culture that reflects the origins, culture, and permissionless nature of the crypto community, end quote. Speaking of crypto, I found this interesting. Chainalysis estimates $40 billion worth of illicit crypto transactions took place in 2024. But the Interesting bit is that criminals seem to have largely shifted from mostly using Bitcoin in 2021 to mostly using stable coins now. Quoting CoinDesk, criminals are changing their habits when it comes to sending illicit funds in 2021, around 70% of all illicit transactions involved Bitcoin. That has now flipped to be dominated by stable coins. BTC now accounts for around 20% of all

illicit transactions, whilst stable coins occupy the majority with 63%. Privacy coin Manero is also a notable inclusion on the list due to its prevalence on darknet markets. Altcoins account for around 10% of all illicit transactions in total, end quote. Meta says it has fired roughly 20 employees who leaked confidential information outside the company and expects there will be more. I don't usually do stories like this, but I thought it was notable given how leaky meta tends to be as a company. Quoting the Verge, we tell employees when they join the company and we offer periodic reminders that it is against our policies to leak internal information no matter the intent meta spokesperson Dave Arnold tells the verge exclusively. We recently conducted an investigation that resulted in roughly

20 employees being terminated for sharing confidential information outside the company, and we expect there will be more. We take this seriously and we'll continue to take action when we identify leaks. Meta has ramped up its efforts to find leakers due to a recent influx of stories detailing unannounced product plans and internal meetings, including a recent all-hands led by CEO Mark Zuckerberg. After we and other outlets reported on what Zuckerberg said during that meeting, employees were warned not to leak. In comments that were subsequently leaked, CTO Andrew Bosworth then told them that the company was making progress on catching people. Morale inside meta has suffered since Zuckerberg announced drastic changes to content moderation policies, ended company DEI programs, and said recent layoffs targeted low performers.

Thursday's public announcement about firing employees doesn't provide any information about what material they leaked, who the employees are, or who they leaked to. There's a funny thing that's happening with these leaks, Bosworth said during an internal meeting in early February. When things leak, I think a lot of times people think, ah, okay, this is leaked, therefore it'll put pressure on us to change things. The opposite is more likely, end quote. This is actually bad news for this podcast, because I guess there will be less meta news if we don't hear from these leakers. Time for the weekend long read suggestions. Today was an unusually newsy Friday, so only room for one. a bit of a local interest story. Wired takes a look at how the New York City subway is using

Google Pixel phones to identify 92% of defect locations on subway tracks, later found and followed up by human inspectors. Quote, between September and January 6 Google Pixel smartphones hitched free rides on four New York City subway cars. Specifically, they took the A train as it ping ponged the 32 miles between the northern tip of Manhattan and the southern reaches of Queens. The phones weren't stowaways or strays, and an extremely sharp-eyed passenger could tell because they were inside plastic enclosures and secured via brackets to the car's undersides and interiors. While people inside the cars use their smartphones to write emails or scroll Instagram or explore Roblox, the subway operators were using these phones sensors, accelerometers,

magnometers, and gyroscopes, and for those attached to the car's exterior, additional external microphones, to listen intently. The phones were part of a brief experiment. by New York City's Metropolitan Transportation Authority and Google into whether cheap, mostly off-the-shelf technology could supplement the agency's track inspection work. Google Public Sector, the division that undertook the work, didn't charge the MTA for this initial experiment. Today, inspections are carried out by human inspectors who, together, walk all 665 miles of New York City subway tracks, eyes peeled for problems like broken rails, busted signals, and water damage.

Thrice annual rides by specialized sensor-laden train geometry cars also capture and upload more sophisticated data on the status of the city's rail infrastructure. New York City transits work with the experimental technology, which Google calls track inspect, suggests that audio, vibration and location data, collected relatively cheaply and used to train artificial intelligence prediction models, can supplement that inspection work. It can point humans towards suspicious rattles, bangs, and squeals, suggesting what kinds of tools they'll need to make the repairs before they get there. Throughout the four-month project, the tech was able to identify 92 percent of

defect locations later pinpointed by human track inspectors, the MTA said. Eventually, the tech could become a way we could minimize the amount of work that's done to identify these defects and point inspectors in the right direction so they can spend time fixing instead of identifying and go directly there and do the work, says Demetrius Cricklaw, the agency's president. In the future, the MTA hopes to create a modernized system that automatically identifies and organizes fixes for track issues. For the system's 3.7 million daily riders catching defects before they become problem, could be the difference between getting to work or school on time and getting mired in

in unexpected delays. The goal with this project is to find issues before they become a major issue in terms of service, said Cricklaw. The collaboration with Google will now expand to a full pilot project. The MTA says where Google will build a production version of the tech and put it in the hands of track inspectors themselves, end quote. Nothing more for you today. Nothing more for you this week.

Nothing more for you this month. No weekend bonus episode this week. Talk to you on Monday.