Tech Brew Ride Home - Thu. 02/13 – Do We Maybe Have A Roadmap For GPT-5?

Episode Date: February 13, 2025

We maybe have a roadmap for GPT-5, but also, is it not really GPT-5 just a renaming of what they already have in the pipeline? It’s kind of weird. Elon gives his conditions for dropping his takeover bid. Utility companies say the AI hype is real, but what if, in the end, the end users don’t actually show up? Sponsors: Shopify.com/ride Links: OpenAI postpones its o3 AI model in favor of a ‘unified’ next-gen release (TechCrunch) Elon Musk will withdraw bid for OpenAI’s nonprofit if its board agrees to terms (TechCrunch) Musk, Altman Spar Over $97.4 Billion OpenAI Bid in Court Filings (Bloomberg) The Apple TV app is now available on Android: watch Apple TV+ and MLS Season Pass (9to5Mac) Meta Opens Facebook Marketplace to Rivals in EU Antitrust Clash (Bloomberg) Data Center Power Demand Almost Doubled in Virginia, Utility Says (Bloomberg) AI Agents Are Everywhere…and Nowhere (WSJ) 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 Thursday, February 13th, 2025. I'm Brian McCullough today. We maybe have a roadmap for GPT5, but also, is it not really GPT5? Just a renaming of what they already have in the pipeline. It's kind of weird. Elon gives his conditions for dropping his takeover bid. Utility companies say the AI hype is real.

But what if, in the end, the end users don't actually show up. Here's what you miss today in the world of tech. The whole Elon bid thing has sucked up all the attention, but in the background, this is a lot of has quietly been an important moment for Open AI. They need to raise more money, thus the wooing of Masassan, but also they need to demonstrate that they're still leading in the AI race, and that was true even before the whole Deep Seek thing happened. Thus, the 12 days of announces in December, remember? So a bunch of announces from San Altman and Company over the last day. Basically, they were giving the outlines of their roadmap going forward into GPT-5. which will apparently include O3, which is no longer set to ship as a standalone model.

Also, GPT4.5 will apparently be OpenAI's last non-chain-of-thought model, quoting TechCrunch. In a post on X on Wednesday, Altman said that in the coming months, open AI will release a model called GPT5 that, quote, integrates a lot of OpenAI's technology, including O3 in its AI-powered chatbot platform, chat GPT, and API. As a result of that roadmap decision, OpenAI, no longer plans to launch O3 as a standalone model. The company originally said in December that it aimed to release O3 sometime early this year. Just a few weeks ago, Kevin Weill, Open AI's chief product officer, said in an interview that O3 was on track for a February March launch. Altman also announced that OpenAI plans to offer unlimited chat access to

GPT5 at the standard intelligence setting, subject to abuse thresholds. Once the model is generally available. Altman declined to provide more detail on what this setting and these abuse thresholds entail. Subscribers to ChatGPT Plus will be able to run GPT5 at a higher level of intelligence, Altman said, while ChatGPT Pro subscribers will be able to run GPT 5 at an even higher level of intelligence. GPT will incorporate voice, canvas search, deep research, and more, he added, referring to a range of features OpenAI has launched in ChatGBT over the past few months. A top goal for us is to unify our models by creating systems that can. can use all our tools, know when to think for a long time or not, and generally be useful for

a very wide range of tasks. Before GPT-5 rolls out, OpenAI plans to release GPD4.5, a model coden named Orion in the next several weeks, according to Altman. Altman says this will be the company's last non-chain-of-thought model. Unlike O3 and OpenAI's other reasoning models, non-chain-of-thought models tend to be less reliable in domains like math and physics. It seems that OpenAI is fully embracing the reasoning model trend it arguably kick-started with its first reasoning model 01 late last year. Reasoning models effectively fact-checked themselves, which helps them to avoid some of the pitfalls that normally trip up models. This fact-checking process incurs some latency. Reasoning models take a little longer, usually seconds to minutes longer to arrive at solutions,

but they tend to be both more reliable and capable. In recent social media posts, Alden admitted that DeepSeek has lessened OpenAI's technological lead in AI and said that OpenAI would pull up some releases to better compete, end quote. Ethan Mollick tweeted, is not unexpected. Reasoners becoming the dominant form of LLM after GBT 4.5, but it is good to have some transparency and commitment to free frontier models for all deserves real applause. Still would be nice to have a lab be fully transparent about their one to three year roadmaps, end quote. Though Gary Marcus tweeted this, quote, Sam Altman tried today to put a positive spin on what looks like a major retreat. Per at Wall Street Journal reporting in December,

Project Orion, which aimed to be GPD 5, was behind schedule and running up major bills. Today, Altman recristened Orion as GPD 4.5 rather than 5, which likely means that despite major bills, Orion never met the expectations so-called scaling laws had implied. AI will soldier on, perhaps with new inventions, but pure scaling of data and compute didn't even get us to GPT5, let alone GPT6. Even with the discovery of test-time compute, GPD5 remains elusive, end quote. And Fergal Reed tweeted this, Low confidence hot take, but I read this as bearish for OpenAI,

and indeed for anyone trying to massively scale pre-training. The implication of GPT5 being powered by 03 and not 4.5, which I presume is a scaled-up model, nor a 4.5 that does reasoning, again presuming a large 4.5, indicates that the last 12 to 18 months of attempting the scale pre-training has probably not resulted in another 10x jump for pre-trained models, especially when you read this with Gemini 2's benchmark performance. Yes, test time compute is scaling well, that's real.

Maybe pre-training has scaled okay logarithmically, but test-time is just doing so well. Smaller model plus test time just economically beats large pre-training runs. But given deep-seek R1, S-1, etc., converting base models to reasoning models isn't that expensive, so that doesn't seem to justify the capital structure of something like, opening eye, end quote. Also, just an FYI that opening AI plans to give free chat GPT users unlimited GPD5 access at the standard intelligence setting, as I said earlier, plus and pro users will get to run GPD5 at higher levels, quoting a gadget. We want AI to just work for you.

We realize how complicated our model and product offerings have gotten, Altman wrote. We hate the model picker as much as you do and want to return to Magic Unified Intelligence. In its current iteration, forcing chat GPT to use a specific model, such as O3 Mini involves either tapping the reason button in the prompt bar or one of the options present in the model picker, which appears after the chatbot answers a question. If you pay for chat GPT Plus or Pro, that drop-down menu can get pretty long with multiple models and intelligence settings to choose from. Once GPT-5 arrives, open-eye plans to offer free users unlimited access to the model subject to abuse thresholds at the standard intelligence setting. Altman did not

provide an exact timeline for either GPT 4.5 or GPT5, other than to say they could arrive with within weeks or months, end quote. But back to the whole Elon bid business. According to a court filing, Elon Musk will apparently withdraw his $97.4 billion bid if OpenAI's board of directors halt its conversion to a for-profit company. Quoting TechCrunch, should the charity's assets proceed to sale, a Musk-led consortium has submitted a serious offer that would go to the charity in furtherance of its mission, the filing reads. However, if Open AI's board is prepared, to preserve the charity's mission and stipulate to take the for-sale sign off its assets by halting its conversion, Musk will withdraw the bid, end quote. The filing submitted to the U.S.

District Court for the Northern District of California claims that Musk's offer to buy OpenAI's nonprofit is serious and that the nonprofit must be compensated by what an arm's length buyer will pay for its assets. In a filing earlier on Wednesday, attorneys for Open AI called Musk's move to take control of the company an improper bid to undermine a competitor and a contradiction of his position in court that a transfer of the startup's assets through restructuring would breach its mission as a charitable trust, end quote. Indeed, in another filing, OpenAI says Elon Musk's bid contradicts his attacks in his lawsuit against Open AI that its assets can't be transferred away for private gain. This is a filing in that original case where he originally sued to halt the

move for the for-profit plan. So Musk was offering to drop. the takeover bid, I guess, to stop this case if they stopped going for profit, quoting Bloomberg. Musk's move to take control of Open AI contradicts his position in court that a transfer of the startup's assets through restructuring would breach its mission as a charitable trust, according to Wednesday's filing. Out of court, those constraints evidently do not apply so long as Musk and his allies are the buyers. The Open AI lawyers said Musk would have OpenAI transfer all of its assets to him for his economic benefit and that of his competing AI business and handpicked private investors, end quote.

The judge in the case U.S. District Judge Yvonne Gonzalez-Rogers signaled at a February 4th hearing that she wasn't convinced she needs to take immediate action against OpenAI. She said she will probably let Musk take Open AI to trial and require him to testify over at least some of his claims. Lawyers told the judge the earliest a trial could take place is in late 2026, end quote. Apple has released an Apple TV app for Android phones and tablets limited to TV Plus, MLS season pass and MLB Friday Night Baseball, and the app doesn't support casting. But clearly this is a play to get subscribers outside of the Apple ecosystem.

Quoting 9 to 5 Mac. The app lets users log in with their existing Apple account or create a new account through the app and subscribe using Google Play billing. The app is rolling out now, supporting Android 10 or newer devices. The app supports the fundamental Apple TV app features users will be familiar with from the app on Apple devices, including the continue watching queue, offline downloads, and search. Playback progress sinks across all your devices so you can start watching a show on your TV and then continue in bed on your Android phone, for instance. However, the app is limited to Apple TV Plus, MLS season pass, and MLB Friday Night Baseball content.

iTunes store purchases and rentals are not shown in the app at all, nor can users access their previously purchased library through the Android app. In some ways, though, this makes for a cleaner experience with a simple layout of four tabs on the phone app, Apple TV Plus, MLS downloads, and search. On tablet form factors, this interface is presented in a floating sidebar layout. Rather than a direct port of the iOS app, Apple is using native Android UI components where applicable, such as context menus when long pressing on an item. However, this is still a 1.0 release, and there are some notably absent features.

The Android app does not support notifications or casting, for instance. An app for Google TV living room devices has existed for many years, but this is the first time the Apple TV app is natively available for Android phones and tablets. Those customers wishing to enjoy Apple originals on non-Apple mobile devices were previously limited to a mediocre web app experience available in the browser. Additionally, starting today, the Google TV Apple TV app will also allow customers to subscribe directly to TV Plus and MLS season pass using Google Play billing. This solves another pain point as previously the app forced users to first make an account

and subscribe on another device. Today's launch is both symbolic and significant and perhaps the biggest statement yet that Apple cares about Apple TV Plus becoming successful as a standalone service rather than a mere ecosystem add-on, end quote. Meta says it will open up Facebook Marketplace and let classified ad firms in Europe pay to place their listings there. This comes after a pilot with eBay and is in direct response to the EU's recent 798 million Euro fine. Quoting Bloomberg. In a statement, Facebook said that despite the changes, it continues to

disagree with the European Commission's decision last year, which saw Meta punished for abusing its dominance. The company also reiterated earlier comments from Meta Chief Mark Zuckerberg in calling the EU antitrust fines tantamount to a tariff regime. The Brussels-based commission said it is currently assessing whether Meta has fully complied with the decision, end quote. European Union antitrust regulators issued a directive in November requiring Meta to separate its Facebook. Marketplace classified service from its main social network, and to stop imposing what it deemed anti-competitive conditions on competing online marketplaces for used goods. The decision, which included financial penalties, marked one of the last major enforcement actions by outgoing EU

competition commissioner Margith Vestager. During her tenure, the Brussels regulator imposed multiple antitrust fines totaling over 8 billion euros against U.S. tech firms with Google Parent Alphabet bearing the brunt of these sanctions. British competition authorities, had also investigated similar concerns regarding Facebook marketplace, but ultimately closed their probe after META agreed to certain commitments, according to the UK's competition and markets authority. You know how I have that list of all-time highs and 52-week-high stocks to give me a heads-up when tech companies suddenly start killing it before the headlines actually reach us? But you know what has been near the top of those lists recently?

Boring old utility companies. Why? Well, AI, of course. As we've discussed, demand for electricity is through the roof, so naturally, business is booming if you produce electricity. Case in point, one of the stocks making all-time highs is Dominion Energy, which now says data center power demand in Virginia, which is home to the so-called data center alley, almost doubled in the second half of 2024, rising to 40.2 gigawatts in December 2020, alone. That's up from 21.4 gigawatts just back in July. Yeah, that'll pop your stock. Quoting Bloomberg.

Demand for power is surging with the development of data centers and artificial intelligence, along with manufacturing and the increasing electrification of the economy. Northern Virginia, which has the biggest concentration of the facilities in the world, has earned the nickname of Data Center Alley. Dominion still expects big demand growth even after Chinese AI company Deepseek upended some of those expectations last month when it released a model that appeared to be much more energy efficient. What's undeniable is that data center, growth in Virginia is not slowing down. In fact, it's accelerating. Dominion Chief Executive

Officer Bob Blue said on the call. Blue acknowledged the spike in request was likely boosted by a system the company instituted in August to evaluate new requests for power in batches in the order they're received. Developers reimburse Dominion for costs. One gigawatt is roughly the output of a nuclear reactor and can power about 750,000 homes, end quote. Yeah, but again, what if the demand for this stuff never actually materializes on the end-user side. A survey of attendees at the Wall Street Journal's CIO Network Summit found that while 61% of chief information officers reported experimenting with AI agents, 21% are not using them at all, citing a lack of reliability. Quote, that's in stark contrast to the vendors selling this technology who say it will be too late

for businesses to wait for all of the technologies kinks to be ironed out. Vendors like OpenAI, Microsoft and Sierra are banking on the fact that enterprises will be ready sooner rather than later to take on new workforces of AI agents that automate away much of the daily toil for their employees. Except that it is imperfect, Brett Taylor, co-founder and CEO of the Agentic AI startup Sierra and chairman of OpenAI said at the summit, rather than say, will AI do something wrong? Say, when it does something wrong, what are the operational mitigations that we've put in place to deal with it? Not all business technology leaders are ready to take on that risk. 29% of attendees at the summit said cybersecurity and data privacy is their primary concern around

using AI agents. Jared Spitaro, corporate vice president of AI at work at Microsoft, said at the summit that the business software giants plan to help companies get value from AI involves combining its co-pilot AI assistant with AI agents. When you use that simple frame and assistant, copilot plus agents, we think you can wind up with the right pattern, he said. Still, the promise of AI agents and AI largely could be even. further away than a few years in the future. 75% of summit attendees polled, said they believe AI is currently driving a small amount of value for their investments, but not enough.

Sometimes, however, companies are in the mode of having hammers looking for nails, said Jim Sider's chief information officer of data analytics firm Palantir, also speaking at the summit. They're so scared to be left out of a revolution in progress that they buy an AI thing of some kind and they try to figure out how to drive value with it, Sider said, end quote. Okay, so the 2000th episode of this podcast is happening later this month, and I want to do something special for it, like probably a listener call-in episode. We did that a couple of times years ago, and Riverside supposedly has a thing that allows people to watch live and be brought up from the audience, so I want to give that a test run. Anyone want to help me kick the tires on that? I know that this is a bit sudden, but watch my Twitter and Blue Sky account today at around 2.30 Eastern.

If I can put a link in there, we'll do a quick test stream. I won't publish it, but if you want to participate, it will help me see how the system works before we record the episode proper. If I don't do it at 2.30 Eastern today, I'll do it at 1 p.m. Eastern tomorrow, so check my socials or email me at Brian at Techmeme.com if you want to participate, and maybe I'll just email you the link. Thanks in advance if you'd like to help. Again, this is a test run, and it won't be published, but it is going to be live.

So keep that in mind if you join in. Talk to you maybe in a couple hours.