Tech Brew Ride Home - Wed. 03/26 – The Day Of AI Announces

Episode Date: March 26, 2025

It’s all AI today basically. New Gemini “thinking” models. New “deep reasoning” agents for Copilot. But the really big news is the new image generator from OpenAI. Fidelity wants to get in the stablecoin business. And if Europe wants to create its own Starlink, it’s got some serious hurdles to overcome. Links: OpenAI rolls out image generation powered by GPT-4o to ChatGPT (The Verge) Gemini 2.5 Pro is Google’s ‘most intelligent AI model’ with thinking built-in (9to5Google) Microsoft adds ‘deep reasoning’ Copilot AI for research and data analysis (The Verge) Fidelity plans to launch stablecoin in digital assets push (FT) ‘No substitute’: Europe’s battle to break Elon Musk’s stranglehold on the skies (FT) 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 Wednesday, March 26, 2025. I'm Brian McCullough today. It's all AI today, basically. New Gemini thinking models, new deep reasoning agents for co-pilot. But the really big news is the new image generator from OpenAI. Fidelity wants to get in on the stable coin business. And if Europe wants to create its own Starlink, it's got some serious hurdles to overcome. Here's what you miss today in the world of tech. Okay, as I said, this is going to be a day of AI model announces. There are a bunch of them. But there's a They're all pretty significant. First up, OpenAI has rolled out GPT40 image generation natively inside chat GPT across all subscription tiers and says the Omni Model model is a, quote, step change above previous models. And I've got to say, I've been playing around with it all morning and it's, yeah, it's definitely very good. So good, in fact, that I'm already considering dropping my Mid Journey subscription and my Claude subscription and going back. exclusively to chat GPT. Check my socials on Blue Sky or Twitter for my Jack Dorsey image to see why I'm so

wowed. I'm Brian MCC on Twitter and at Brian MC, I think, on Blue Sky, quoting the verge. Open AI is integrating new image generation capabilities directly into chat GPT starting today. This feature is dubbed images in chat GPT. Users can now use GBT 4O to generate images within chat GPT itself. This initial release focuses solely on image creation and will be available across chat GPT Plus, pro, team, and free subscription tiers. The free tiers usage limit is the same as Dali. Spokesperson Taya Christensen told the Verge, but added that they didn't have a specific number to share, and these may change over time based on demand.

Per the chat GPT FAQ, free users were previously able to generate three images per day with Dali 3. As for the fate of Dali, Christensen said, fans will still have access via a custom GPT. This model is a step change above previous models. Research Lead Gabriel Goh told The Verge, adding that the team use the GPT40 Omni model or a model that can generate any kind of data like text, image, audio, and video foundation for this feature.

Some of the improvements Goh noted include binding, which refers to how well AI image generators maintain correct relationships between attributes and objects. A model with poor binding, for instance, might get a prompt for a blue star plus a red triangle and create a red star and no triangle. Most image models struggle with this, Go said, often mixing up colors and shapes when asked to render multiple items, typically around five to eight. He says this new image generation tool can correctly bind attributes for 15 to 20 objects without confusion representing a significant

improvement in accuracy and reliability. Users will also notice an improvement in text rendering, which makes it easier to generate coherent text without typos on an image. In existing tools, you'll often notice that text gets garbled pretty easily. Getting text rendering right was a significant challenge, Goh said. If small titles or text elements have typos or errors, the entire image can become unusable. This was just like a process of iteration that took many, many months to get right, Go said.

While not perfect, he said, the team reached a point where the text quality is consistently usable. Where it tends to blunder is really small text. It's been just many months of small improvements, he said. The system uses an auto-regulatory, aggressive approach, generating images sequentially from left to right and top to bottom, similar to how text is written, rather than the diffusion model technique used by most image generators like Dali that create the entire image at once. Go speculates that this technical difference

could be what gives images in ChatGBTGBT better text rendering and binding capabilities. In a briefing before the feature launch, the team demonstrated several examples showing the system's capabilities, including scientific diagrams like Newton's Prism experiment, with correctly labeled components, multi-panel comics with consistent characters and text bubbles, and informational posters with accurate text. They also highlighted practical applications like creating transparent background images for stickers,

restaurant menus, and logos. If I go to draw an image, I do so with the limitation of my own skill, but also with all of the knowledge of the world that I've built up. ChatGPT Multimodal Product Lead Jackie Shannon explained. The model brings world knowledge to the equation, so when you ask for an image of Newton's Prism experiment, you don't have to explain what that is to get an image back. The new system does take longer to generate images than before, though opening I suggest this is a worthwhile tradeoff.

While we certainly have room to improve on latency, the quality of these images, the capability, the world knowledge really makes up for the additional seconds that they'll spend waiting, Shannon said, end quote. Yes, in my experience, it is pretty slow, but click through on the story link to see those images of that correctly labeled diet. diagram for the Newton Prism experiment. It'll tell you why it's so wowing to a lot of us. I did similar things like creating a diagram of the cross-section of the Earth's core, labeling the various sections for, you know, like a poster you could put up in a classroom, and it got pretty close. I tried to get it to do a map of the D-Day landings on the Normandy beaches, and it kind of failed, but it did have the names of the various beaches right and a decent map of the Normandy Peninsula, even if it got the labeling wrong. Seriously, it's that level of stuff

that wasn't possible before. Again, check out my socials. Jack Dorsey in the style of a Wes Anderson movie poster. It even got the friggin' font right. Next, Google has debuted Gemini 2.5 thinking models for developers and Gemini Advanced Subscribers, starting with Gemini 2.5 Pro experimental, which tops some benchmarks. Quoting 9 to 5.5, of Google. Notably, all the models in the Gemini 2.5 family, including future ones, are thinking models capable of reasoning through their thoughts before responding, resulting in enhanced performance, and improved accuracy. According to Google, Google is, quote, building these thinking capabilities directly into all of its models to allow them to handle

more complex problems and support even more capable context-aware agents. Compared to 2.0 flash thinking, which was first revealed in December and got an update this month, Google is no longer explicitly attaching the thinking label. Users can show thinking in the Gemini app to see the train of thought. Gemini 2.5 features a new level of performance by combining a significantly enhanced base model with improved post-training. Gemini 2.5 Pro, Gemini 2.5 Pro XP.0325 and codenamed Nebula, is the first model in this family.

Aimed at complex tasks, Google notes how it tops the L-M-A-L-Rena leaderboard, which measures human preferences by a significant margin. It also leads on math and science benchmarks without test time techniques that increase costs like majority voting. This is quoting Google. It also scores a state-of-the-art 18.8% across models without tool use on humanity's last exam, a dataset designed by hundreds of subject matter experts to capture the human frontier of knowledge and reasoning. In addition to native multimodality, Gemini 2.5 Pro has a 1 million token context, window with $2 million coming soon. Gemini 2.5 Pro Experimental is rolling out first to Gemini

Advanced and Google AI Studio with Vertex AI following in the coming weeks, end quote. Not to miss the party, Microsoft unveiled AI deep reasoning agents for 365 co-pilot researcher based on OpenAI's deep research model and analyst based on O3Mini coming in April. Quoting the Verge, researcher relies on OpenAI's deep research AI model to pull off complex multi-step research, along with access to third-party data via connectors to sources like Salesforce or ServiceNow, so that business customers can derive insights from across their tools. Analyst is based on the 03 mini reasoning model from OpenAI, and Microsoft claims that with chain of thought reasoning, it's capable of turning raw data into spreadsheets, running

Python code that you can view while it's running, and operate on the level of a skilled data scientists pulling together reports. Those tools are scheduled to start rolling out in April to Microsoft 365 copilot license holders in an early access program, along with new autonomous agent capabilities that are starting to roll out now in Copilot Studio. Microsoft claims the new agent flows in Copilot are powerful enough to automate any task you can imagine with rule-based workflows that include AI actions. The LinkedIn announcement describes situations like an agent flow that directs feedback emails to the correct team, but we'll have to see it in action to find out how that's better

than adding a checkbox or two or how well its low-code experience really works, and if it can deliver on the promises AI companies are making about agents, end quote. Something, something. Everybody wants in on stablecoins. sources tell the FT that Fidelity Investments is in the advanced stages of testing its own stablecoin managed by its digital assets arm. You can't get more mainstream Main Street finance and investing than Fidelity, so if they're in, quote, the Boston-based fund. firm is in the advanced stages of testing its token, which is designed to act as cash in cryptocurrency markets and will be managed through its digital assets arm, according to two people with knowledge

of the plans. Fidelity's planned launch is part of its expansion into the nascent market for tokenized versions of U.S. Treasuries, having been involved in digital assets for more than a decade. Late last week, it filed to launch a digital version of a U.S. money market fund at the end of May in direct competition with traditional asset-manage rivals BlackRock and Franklin Templeton. Its move comes as Washington begins sweeping changes in the oversight of cryptocurrencies following the election of President Donald Trump. Trump has pledged to promote the growth of lawful and legitimate dollar-backed stable coins to support U.S. currency and called for supporting legislation to be ready to be signed into law by August. Politicians in Washington are debating rival bills to begin regulation of stablecoins, which are designed to hold a constant value per coin and serve as ready cash reserves outside the regulated banking system. Advocates see tokenized money market funds, in contrast to stablecoins,

as regulated onshore deposits. However, critics have argued that existing tokenized funds lack the deep liquid secondary markets that stable coins typically have. Tokenization could transform the financial services industry, according to Cynthia Loebassette, head of digital asset management at Fidelity Investments. One use case to improve the efficiency of capital markets was using tokenized assets as collateral to meet margin requirements in trading, she added, end quote. I guess I need to open a file called something like the is Europe decoupling its tech stack from Silicon Valley file, because here's another example. The FT reports that as the EU considers funding a Starlink rival, experts say no single European

network can match Starlink's range right now, likely requiring a patchwork of satellites. Quote, the prospect of a new European push for space sovereignty has boosted shares in heavily indebted operators such as UTELSAT and SES in recent weeks. But even with EU funding's success, will not come easy. No single European network can replicate such a wide variety of applications, say industry experts. Instead, a European solution would be made up of a patchwork of satellites in different orbits, offering differing performances and requiring different user terminals for different networks. Today, there is no substitute for Starlink, said Panduro, but there may be alternatives that without being a substitute can help to alleviate the absence of those capabilities.

A senior executive at a rival satellite operator put it more bluntly, Starlink is so disruptive, so cheap, so pervasive, and so excellent. At the heart of the problem lies the failure of legacy operators in Europe and, more widely, to match the agility of Starlink. Many assumed the challenges of operating in space would protect them from disruption while they tried to offset declining broadcast revenues with new connectivity businesses. But when Starlink arrived, helped by cheap launch services from parent SpaceX, everything changed completely, said Jean-Baptiste Theo Poe, principal at space consultancy Nova Space. years, Musk's broadband service has launched 7,000 satellites into the relatively

unexploited region of space known as low Earth orbit, or LEO. It is now the world's biggest satellite operator and has won broadband contracts with airlines, shipping groups, and governments. Traditional operators, such as SES and UTELSAT, had hoped those sectors would give a new lease of life to their satellites in higher geostationary or GEO orbits of about 36,000 kilometers above the Earth. Flying at lower altitudes, roughly 550,000, kilometers in Starlink's case requires thousands of satellites and billions in investment for a global service. But by operating in LEO, Starlink has delivered lower latency, the time it takes for a signal to travel from Earth to a satellite and back, and higher speed connectivity to the

mass market than many legacy operators. Its dense network of satellites had proved to be inherently resilient, said James Trevelyan, Executive Vice President at Speedcast, which sells capacity on both Starlink and One Web networks. Starlink was also designed for the consumer market and heavily subsidized its $2,000 terminals early on, selling them at $500 to $600. In light of the disruption and price pressure, legacy operators have focused on higher-performance geo-networks and new orbits. In 2022, U-TELSat announced its acquisition of one web, which operates in LEO like Starlink, but at an altitude of 1,200 kilometers. SES, meanwhile, has expanded in Medium Earth orbit or MEO with its O3BM-Power network

and is acquiring rival geo-operator IntelSat. Both companies took on substantial debt to fund their multi-orbit strategies, moves that have yet to bear fruit. OneWeb has struggled to meet initial revenue targets due to slower than expected ground station rollouts. Analysts also believe that heavily indebted UTELSAT will struggle on its own to fund a new generation of OneWeb satellites, widely regarded as necessary to be competitive with Starlink's more capable technology. Europe's flagship 10 billion euro iris 2 project, which aims to provide secure government communications from 2030 and is its most ambitious space program in a decade, will be critical to unlocking funding for this. SES, meanwhile, suffered power problems on the first satellites in its Mpower 03B generation in 2023, resulting in extra investment and impairment charges. It is impossible to replace Starlink in a day, SES chief executive Adele Al-Ale, told the

the Financial Times. However, in the longer term, Europe could do so, he added, a single-orbit network is not resilient enough. You need multi-orbit for resilience for backup for being able to move traffic around. The stakes are now higher as policymakers focus on Europe's patchwork industry. The new landscape repositions UTELSAT and SES as vital components of sovereign defense infrastructure, said Alexander Petrick, head of small and mid-cap research at Bernstein. One web terminals designed for business and government rather than consumers are also bulkier, more complex to configure initially, and cost about $5,000 to $10,000 each. The brutal reality is that the terminals remain a big blocker for a European alternative,

said Speedcast Trevelyan. Those who know the One Web Network say it could still work. Smaller terminals are coming to market. The Commission's desire for a European alternative to Starlink presented an unexpected opportunity, but it will not be easy and it won't be enough to resolve questions about the long-term future of Europe's satellite operators, said, Pierre Lionette, search director at the Trade Body ASD Erospace. Amazon's planned LEO Constellation Project Kuiper will further disrupt pricing and competition

when it is eventually operational. When that comes, it could be a real problem for European companies because at the end of the day, their current technology is inferior, Trevelyan said. Short sellers are estimated to have lost about $150 million on the recent surge in UTELSAT and SES shares over the past month, according to S3 partners, but analysts question whether the rally is really sustainable. Starlink has deep pockets and a vertical integration advantage, said Nicholas Kordowski, head of non-financial fixed income research at Aberdeen, and that really hasn't gone away, he said,

end quote. Finally, just to note this, Apple announced WWDC 2025 for June 9th through the 13th, which will be an entirely online event free for developers with an in-person special event at Apple Park on June 9th. So mark your calendars. Talk to you tomorrow.