Tech Brew Ride Home - Tue. 04/15 – Why Is OpenAI Going Backwards (Name-Wise)?

Episode Date: April 15, 2025

OpenAI releases its latest next gen models but you wouldn’t know it by the nomenclature because the numbers are going backwards. What’s up with that? Apple is tying itself in a pretzel to train on user data but still stick to privacy. And a big rundown of the first day of the big Meta antitrust trial. Sponsors: MackWeldon.com promocode BRIAN Links: OpenAI’s new GPT-4.1 AI models focus on coding (TechCrunch) OpenAI launches another model before GPT 5 — here’s what this one can do (Tom's Guide) Nvidia says it plans to manufacture some AI chips in the US (TechCrunch) Apple to Analyze User Data on Devices to Bolster AI Technology (Bloomberg) Zuckerberg testifies as FTC, Meta trade opening salvos in antitrust trial (Politico) Mark Zuckerberg Takes Stand to Defend Meta Against Antitrust Suit (NYTimes) 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 Tuesday, April 15th, 2025. I'm Brian McCullough today. OpenAI releases its latest next-gen models, but you wouldn't know it by the nomenclature because the numbers are going backwards. What's up with that? Apple is tying itself in a pretzel to train on user data, but still stick to privacy and a big rundown of the first day of the big meta-antitrust trial. Here's what you miss today in the world of tech. OpenAI has released GPT 4.1, GPT 4.1 Mini, and GPT. G.T4.1 Nano, which they claim Excel at coding, instruction following, and long context understanding available via its API. Okay, now obviously, I'm about to quote from a bunch of places telling you

what is new and different in these models, but you may be thinking, A, Brian, another new model, can it really be that different enough for me to care? And B, wait, I thought we already had GPT4.5. Are we going backwards? Hang on. We'll address both. both of those issues in a second. First, quoting TechCrunch. OpenAI on Monday launched a new family of models called GBT 4.1, yes, 4.1, as if the company's nomenclature wasn't confusing enough already. There's GPD 4.1, GPD 4.1 mini, and GPT 4.1 nano, all of which OpenAI says Excel at coding and instruction following. Available through OpenAI's API, but not chat GPT, the multimodal models have a 1 million token context window, meaning they can take in roughly 750,000 words in one go, longer than the novel war and peace.

It's the goal of many tech giants, including OpenAI to train AI coding models capable of performing complex software engineering tasks. OpenAI's grand ambition is to create an agentic software engineer, as CFO Sarah Fryer put it during a tech summit in London last month. The company asserts its future models will be able to program entire apps end-to-end handling aspects such as quality assurance, bug testing, and documentation writing. GPD 4.1 is a step in this direction. OpenAI claims the full GPT4.1 model outperforms its GPT40 and GPD40 mini models on coding benchmarks. GPT41 Mini and nano are said to be more efficient and faster at the cost of some accuracy, with OpenAI saying GPT41 Nano is its speed. and cheapest model ever. GPT 4.1 costs $2 per million input tokens and $8 per million output tokens. GPT41 Mini is 40 cents per million input tokens and $160 per million output tokens, and GPT41 Nano is 10 cents per million input tokens and 40 cents per million output tokens, end quote.

And quoting Tom's guide, what does this mean for the average person? This side of OpenAI's market is pretty niche in terms of who will be using the models. Coders and researchers will be making full use of the GPD 4.1 series looking to these models to better understand the inner workings of AI and using its brains to accomplish complex coding tasks. However, while this won't directly affect most of us, it does show the development being made by OpenAI. Most noticeably, the ability for its models to take in large amounts of contextual data with higher token limits. Where is GPT5? Open AI has for a long time now been promising the release of GPT5. This would be the next powerhouse behind ChatGPT, and in theory

the biggest update OpenAI has released in a very long time. Addressing the mess that is their naming system recently, Sam Altman, CEO of OpenAI stated on X, quote, how about we fix our model naming by this summer and everyone gets a few more months to make fun of us, which we very much deserve until then, end quote. Aside from accepting the name blame here, it seems Altman is hinting toward a move forward for naming, or in other words, the start of GPT-5. This would line up well with previous hints toward release dates, and as long as there are no delays or surprises, we may see the launch of GPT-5 in the next couple of months, end quote. So, again, to sum up, is this just another iterative release to catch up with the hotness that is vibe coding and

agents? It certainly seems like it, quoting Dan Mack. GPT-4-1 seems like a response to Anthropic and Google rather than a major leap. I get that it's only Monday, and the big drops are likely to come later in the week, but GPT-4-1 scores 52% on Ader Polyglot. Gemini 2.5 is head and shoulders above at 73%, end quote. Then addressing the second issue I mentioned, does the necessity to do that to just be reactive and match competitors, just add confusion around branding and basic product lineup for open AI for sure? Like, it doesn't help that the numbers are going back. quote, Peky McCormick. If GPD 4-1 is much better at coding than GPD 4.5, I cannot wait to see how impressive GPT 3.7 is, end quote. But note, again, the costs are continuing to come down significantly,

quoting our friend Simon Willison. Added this note to my post to help illustrate how absurdly inexpensive these models have got now. I can use GPD 4.1 nano to generate descriptions of 4,000 images for less than a dollar, end quote. But don't think of the norm. because if you're a power user, if you're the type of person Open AI is targeting here, each of the new models released by anyone can potentially be something new that you plug into whatever it is you're doing. Each one is a different flavor. Each one hopefully allows you to do more, quoting Strawberry Man on Twitter. Okay, vibe coding with at WinSurf AI and 4.1 is another level. I feel like a god, end quote. So, what happens is devs plug.

in various models and to say cursor to see what more they can do and how much faster they can do it, and founders plug in the APIs to make their new agents and startups and whatever. That's why every new model is important to some. A new model can open up a vista of new capabilities or suddenly 10x some area you have been working on. Quoting Aidan McLaughlin on X. Heard from some startup engineers that they lost several work hours gawking stupefied after they plugged 4.1 mini slash nano into every previously expensive part of their stack. You can just do GPT40 quality things, 25X cheaper now, end quote. And yes, let's not lose track of the fact that, in a way, this is kicking the can of GPT5 down the road again.

Sam has absolutely set OpenAI up for a drumbeat of questions about where is five, where is five, where is five going into the summer now, quoting Corey Quinn on X. Not sure what's going on with their sprint to V4, then stay there forever approach to branding, end quote. Bunch of chip news for you. Invidia has started making its Blackwell chips at TSM's Phoenix plant and says it aims to make up to $500 billion worth of AI infrastructure in the U.S. in the next four years. Quoting TechCrunch. Also, Nvidia is building supercomputer manufacturing plants in Texas with Foxconn in Houston and Wistron in Dallas. In Arizona, NVIDIA is partnering with Amcor and Spill for Packaging and Testing Operations.

The company added mass production at the Houston and Dallas plants is expected to ramp up in the next 12 to 15 months. And within the next four years, the company aims to produce up to a half a trillion dollars of AI infrastructure in the U.S., end quote. AMD has announced its first two nanometer silicon fabricated on TSM's N2 process set to debut on its sixth-gen epic Venice chip, which is expected to launch in 2026, according to Tom's hardware. TSMC's N2 is the Foundry's first process technology that relies on Gate All Around or GAA nanosheet transistors. The company expects its manufacturing technology to offer either a 24 to 35% reduction in power consumption or a 15% increase in performance at constant voltage, along with a 1.15x boost in transistor density compared to the previous N3-nometer-class generation. These gains are primarily driven by the new type of transistors and the N2-N2 Nanoflex design

technology co-optimization framework. AMD's announcement comes after its arch-rival Intel delayed the release of its next-generation Zeon Clearwater Forest processor made on its 18A manufacturing technology, which is set to rival TSMC's N2 to the first half of next year, end quote. Mark German says that Apple plans to begin on-device privacy-centric analysis of user data, comparing it to synthetic data in order to improve their AI systems, and this is coming as soon as the iOS 18.5 and MacOS 15.5 betas. quote, Apple will begin analyzing data on customers' devices in a bid to improve its artificial intelligence platform, a move designed to safeguard user information while still helping it catch up with

AI rivals. Today, Apple typically trains AI models using synthetic data, information that's meant to mimic real-world inputs without any personal details. But that synthetic information isn't always representative of actual customer data, making it harder for its AI systems to work properly. The new approach will address that problem while ensuring that user data remains on customers' devices and isn't directly used to train AI models. The idea is to help Apple catch up with competitors such as OpenAI and Alphabet, which have fewer privacy restrictions. The technology works like this. It takes the synthetic data that Apple has created and compares it to a recent sample of user emails within the iPhone, iPad, and Mac email app. By using actual

emails to check the fake inputs, Apple can then determine which items within its synthetic data set are the most in line with real-world messages. These insights will help the company, text-related features in its Apple intelligence platforms, such as summaries in notifications, the ability to synthesize thoughts in its writing tools and recaps of user messages. The new system could theoretically improve Apple's models a key step toward becoming a serious competitor in the hot AI space. The company's artificial intelligence team has seen its products lag behind rivals spurring a recent management shakeup for the Siri voice assistant and related efforts. The company will roll out the new system in an upcoming beta version of

iOS and iPadOS 18.5 and macOS 15.5. A second beta test of those upcoming releases was provided to developers earlier on Monday, end quote. Hey, did you know the antitrust case against meta has begun? It has. It started yesterday in the FTCV meta in court. The US argued meta has a monopoly in the personal social networking market, which it claims includes only Instagram, WhatsApp, Snapchat, and Miwi. Quoting Politico, Meta's CEO Mark Zuckerberg became the first witness to testify Monday at the Federal Trade Commission's landmark antitrust trial that seeks to break up his company. Daniel Matheson, the FTC's lead lawyer, spent the first hour trying to pin down Zuckerberg on the core value proposition of the social media giant, suggesting that meta's platforms are

primarily designed to connect users with friends, family, and other people they know in real life. The question is core to the FTC's claim that meta has a monopoly in the personal social networking market, which the agency contends revolves around connections with friends and family. The FTC claims that market consists of just four platforms, meta-owned Instagram and WhatsApp, plus Snapchat and a much smaller app called Miwi. But Zuckerberg didn't take Matheson's bait, claiming at one point that Facebook's feed has turned away from friends and family toward more of a broad discovery entertainment space, his words. The government is expected to question Zuckerberg over emails he and other executives sent about the purchases of Instagram and WhatsApp.

In its opening statements, the government said the Facebook parent company bought up the competing apps to create a monopoly. In one such email, Zuckerberg famously said, Instagram was growing so fast that the company had to buy it for $1 billion. It's an email written by someone who recognized Instagram as a threat and was forced to sacrifice a billion dollars because META could not meet that threat through competition, said the FTC's lead lawyer, Daniel Matheson. If the FTC convinces U.S. District Judge James Boasburg, that Meta's 2012 and 2014 acquisitions of Instagram and WhatsApp were illegal, the agency will try to split up the $1.4 trillion company. A breakup of that size hasn't been attempted since telephone monopoly AT&T was unwound 40 years ago. Meta's opening argument suggests the company will focus overwhelmingly on combating the FTC's definition of the personal social networking market.

Its lawyers led by Kellogg Hanson partner Mark Hanson claimed the government's definition of the market, a key step in establishing if there is a monopoly, excludes rivals like TikTok, YouTube, and I message. That's indefensible. That's gerrymandering, Hansen said, of the FTC's curtailed social media market, end quote. And quoting the times. In a packed courthouse in the U.S. District Court for the District of Columbia, lawyers for the Federal Trade Commission presented Mr. Zuckerberg with a binder full of dated emails and internal communications about his acquisition strategy, pushing him to defend his words. The government has contended that meta illegally cemented a social media monopoly by acquiring Instagram and WhatsApp when they were tiny startups, combining them into the same company, which, was then known as Facebook. I view this all as relatively early thinking, Mr. Zuckerberg said about an email he wrote in February 2012 in which he discussed keeping Instagram going, but not adding more features. In practice, we ended up investing a ton after we acquired it, he said. Mr. Zuckerberg, who is expected to resume testimony on Tuesday, was the first witness in the trial,

Federal Trade Commission v. Meta platforms. Earlier in the day, the FTC opened its first antitrust trial under the Trump administration by arguing that Meta's acquisitions were part of a or bury strategy. Ultimately, the purchases coalesce Mehta's power, depriving consumers of other social networking options and edging out competition, the government said. Meta's lawyers denied the allegations in opening statements, countering that the company faces plenty of competition from TikTok and other social media platforms. The FTC approved the acquisitions of Instagram and WhatsApp more than a decade ago and trying to unwind the mergers would set a dangerous precedent, the lawyers added. The trial poses the most consequential threat to the business empire of Mr. Zuckerberg,

the company's co-founder. If the government succeeds, the FTC is likely to ask META to divest Instagram and WhatsApp, potentially shifting the way that Silicon Valley does business and altering a long pattern in which big tech companies have snapped up younger rivals. Still, legal experts cautioned that it might be challenging for the FTC to win. That's because the government must prove something unknowable. That meta would not have achieved the same success without the acquisitions. It is also extremely rare to try to unwind mergers approved years ago, legal experts said, said. One of the most difficult things for antitrust laws to deal with is when industry leaders purchase small potential competitors, said Gene Kimmelman, a senior official in the Obama

administration's Department of Justice. Meta bought many things that either didn't pan out or were integrated, he added. How are Instagram and WhatsApp different? The efforts continue a year-long bipartisan pursuit to curtail the vast power that a handful of tech companies have over commerce, the exchange of ideas, entertainment, and political discourse, despite attempts by tech executives to court President Trump, his antitress appointees have signaled that they will continue the course. For more than 100 years, American public policy has insisted firms must compete if they want to succeed, said Daniel Matheson, the FTC's lead litigator in the case in his opening remarks. The reason we are here is that meta broke the deal. They decided that competition was too hard and

it would be easier to buy out their rivals than compete with them, he added. presiding over the case is Judge James Boseberg, 62, the chief judge in the federal court. He is already in the national spotlight for rejecting the Trump administration's effort to use a powerful wartime statute to summarily deport Venezuelan migrants it deemed to be members of a violent street gang. Judge Boseberg has said he was never a user of Mena's apps but is familiar with Facebook Live, which has been featured in criminal trials. He took notes as Mr. Matheson explained the government's definitions of social networking and methodology to determine meta was a monopoly. He was equally focused on meta's rebuttal of those definitions.

The FTC argued that Mr. Zuckerberg said in 2016 that Facebook was used to connect actual friends. The agency has argued that meta has had a monopoly in social networking since 2011, and that Snapchat was among the only comparable platforms to Facebook and Instagram. Mr. Zuckerberg described the social media market as much larger than how the government was defining it. Connecting friends and family is, quote, one of the core things the company does, he said, but meta is also involved in the general idea of entertainment and learning about the world and discovering what's going on. Mark Hansen, Meta's lead litigator and partner at the law firm Kellogg Hanson, Todd, Feigel, and Frederick, said meta-face competition from TikTok, LinkedIn, YouTube, and other platforms.

Mr. Hansen said more than half of all engagement on Facebook and Instagram involved videos, which put Meta squarely in competition with TikTok, the fast-growing short video app, end quote. We are in a lovely cabin slash condo in Estes Park, Colorado this week, as I said, for your moment of Zen, here is the sound of the river running right outside my window as I record this this morning. Talk to you tomorrow.