Loading summary
Transcript4 lines
- [00:05]
Brian McCullough
Welcome to the Tech Brew Ride home from Monday, August 11th, 2025 temporarily from London, I'm Brian McCullough. Today Nvidia and AMD are paying the US government 15% of revenue to sell AI chips to China. Remember how with GPT5 they took away the old models? Well, the old models are back. Some critical analysis of the whole GPT5 moment. And you know what is finally never coming back? The AOL dial up Internet. Here's what you missed today in the world of tech. Sources are telling various places that Nvidia and AMD have agreed to pay the US government 15% of revenues from H20 and Mi308 chip sales in China to obtain the export licenses granted to those companies last week. Quoting the FT, the US officials said Nvidia agreed to share 15% of the revenues from H20 chip sales in China and AMD will provide the Same percentage from Mi308 chip revenues. Two people familiar with the arrangement said the Trump administration had not yet determined how to use the money. The Financial Times reported on Friday that the Commerce department started issuing H20 export licenses on Friday, two days after Nvidia Chief Executive Jensen Huang met President Donald Trump. The US Official said the administration had also started issuing licenses for AMD's China chip. The quid pro quo arrangement is unprecedented, according to export control experts. No US Company has ever agreed to pay a portion of their revenues to obtain export licenses. But the deal fits a pattern in the Trump administration, where the president urges companies to take measures such as domestic investments, for example, to prevent the imposition of tariffs in an effort to bring in jobs and revenue to America. AMD did not respond to a request for comment. Nvidia did not deny that it had agreed to the arrangement. It said, we follow rules the US Government sets for our participation in worldwide markets. Bernstein analysts estimate that based on Nvidia's guidance before the controls kicked in earlier this year, it would have sold around 1.5 million H20 chips to China in 2025, generating around $23 billion in revenue. The move follows controversy over the H20 chip. Nvidia tailored the H20 for the Chinese market after President Joe Biden imposed tough export controls on more advanced chips used for artificial intelligence. In April, the Trump administration said it would ban H20 exports to China. However, Trump reversed course in meeting Huang at the White House. Over the following weeks, Nvidia became concerned because the Bureau of Industry and Security, the arm of the Commerce Department that runs export controls, had not issued any licenses. Huang raised the issue with Trump on Wednesday, according to people familiar with the exchange, and BIS started issuing licenses on Friday. The H20 revenue deal comes as Nvidia and the Trump administration face criticism over the decision to sell the chip to China. US security experts say the H20 will help the Chinese military and undermine US strength and artificial intelligence. Beijing must be gloating to see Washington turn export licenses into revenue streams. Liza Tobin, a China expert who served on the National Security Council in the first Trump administration, now at the Jamestown Foundation. What's next? Letting Lockheed Martin sell F35s to China for a 15% commission, end quote? Some BIS officials have also expressed concern about the reversal, according to people familiar with the situation. In a recent letter to Commerce Secretary Howard Letnick, Matt Pottinger, a China expert who was national security adviser in Trump's first term, and 19 other security experts urged the US not to grant H20 licenses. They said the H20 was a potent accelerator of China's frontier AI capabilities and would ultimately be used by the Chinese military. Nvidia said the claims were misguided and rejected the notion that China could use the H20 for military purposes, end quote. And quoting from Bloomberg to call this unusual or unprecedented would be a staggering understatement, said Stephen Olson, a former US Trade negotiator now with the Singapore based ISEA Yousef Isaac Institute. What we are seeing is in effect the monetization of US Trade policy in which US Companies must pay the US Government for permission to export. If that's the case, we've entered into a new and dangerous world. The chip payment arrangement may face legal challenges because it could be construed as an export tax, something that's not allowed under the Constitution. Trade experts said their proposal is the latest direct government intervention into business and finance since Trump returned to the Oval Office in January, as well as a chaotic tariff campaign and persistent criticism of a sitting Federal Reserve chairman. Trump has used his Truth social platform for everything from calling on CEOs to resign to offering commentary on corporate advertising campaigns. The Nvidia and AMD revenue sharing deals may now prompt the White House to target other industries and goods, according to Deborah Elms, head of trade policy at the Heinrich foundation in Singapore. The sky is the limit, she said. You could come up with all sorts of company specific, country specific combinations that would say no one else can trade, but if you pay us directly, then you get the ability to trade. Although Nvidia and AMD agreed to the terms, there are questions about the legality of the agreement, Elm said. The arrangement looks like an export tax, which is again forbidden by the US Constitution. While the US has intervened before, including by taking stakes in private companies after the 2008 financial crisis, a similar deal like the one struck with Nvidia and AMD is hard to remember and without proper oversight could lead to a crony capitalism state, according to Scott Kennedy, senior advisor at the center for Strategic and International Studies in Washington. It represents a huge shift in the way the American economy is supposed to operate, kennedy said. It won't make anyone happy except maybe the Chinese, who will get their chips and watch the US political system go through gyration and domestic tensions. End quote Speaking of CEOs who have been spoken out against by the president, sources also say that Intel CEO Lip Bu Tan is set to meet Trump on Monday to propose ways the US and intel could work together. Quoting the Journal, Tan is expected to have a wide ranging conversation with Trump with the intent of explaining his personal and professional background, the people said. He could also propose ways that the government and intel could work together, they said. Tan hopes to win Trump's approval by showing his commitment to the country and pledging the importance of keeping Intel's manufacturing capabilities as a national security issue, one of the people said. Tan said in a message to intel employees Thursday evening that the US has been his home for more than 40 years and that intel has engaged with the administration to ensure, quote, they have the facts. Over 40 plus years in the industry, I've built relationships around the world and across our diverse ecosystem and I have always operated with the highest legal and ethical standards, he said. End quote. So when I first heard that people were pissed that OpenAI had unified the selector thing so you could only get GPT5 and GPT5 with thinking? When I first heard they were calling for 4.0 and other older models back, I thought it was sort of maybe a joke, but it was not a joke. Quoting PCMag it looks like OpenAI didn't anticipate some users affection for the older GPTs. Everything that made ChatGPT actually useful for my workflow deleted, wrote one Reddit user who particularly misses the GPT 4.0 model. 4.0 wasn't just a tool for me. It helped me through anxiety, depression and some of the darkest periods of my life. It had this warmth and understanding that felt human. The user added on Reddit's forum devoted to ChatGPT, some users are even canceling their paid subscriptions. What kind of corporation deletes a workflow of eight models overnight with no prior warning to their paid users, wrote one subscriber. Other users aren't buying OpenAI's marketing claims about GPT5 being superior and argue the model is a downgrade and a disaster. They have completely ruined ChatGPT. It's slower even without the thinking mode, it has such short replies and it gets some of the most basic things wrong, wrote another user. Well then, OpenAI said ChatGPT Pro users would still be able to select old models for the time being, but still plan to deprecate them in 60 days or so, Sam Altman said. Plus users would be able to keep using GPT 4.0, for example. But now Altman has said OpenAI will bring back GPT 4.0 to ChatGPT, and they are raising reasoning model rates for free and plus users as usage of those reasoning models increases. Quoting Gizmodo On August 7, OpenAI launched GPT5, presenting it as a unified system that would automatically route user queries to the best model for the job. In doing so, it removed the menu that allowed users to choose between older trusted models like GPT 4.0, which was launched in March 2023. For customers paying for subscriptions like ChatGPT plus at $20 a month, the sudden change felt like a betrayal. They had built their professional and creative workflows around a toolkit of different models, each with its own strengths one for creativity, another for pure logic, another for de research. Forcing everyone onto a single new model broke those workflows and removed their ability to cross reference answers to check for errors or hallucinations. The backlash was immediate and fierce, resulting in a cascade of subscription cancellations and online petitions. As the outrage grew over the weekend, Sam Altman took to X to do damage control, acknowledging the company had misjudged the situation and announcing a series of concessions. The biggest news came in a direct reply to a user asking the question on everyone's mind and bringing back 4.0. It's back. Go to Settings and pick show legacy models, altman responded, confirming the return of the fan favorite model. While GPT5 remains the default, users can now opt back into the older versions. Finally, Altman promised more transparency. He announced an upcoming UI change to show users which model is actively responding to their queries, and promised a detailed blog post this week explaining the company's thinking on how we are going to make capacity trade offs. He also shared data showing the immense popularity of the new reasoning models, revealing that daily usage among plus users has jumped from 7% to 24%, subtly justifying the company's initial focus on the new tech. Even if the rollout was a disaster, Reasoning is the AI's ability to think step by step to resolve complex problems.
- [10:49]
Indeed Sponsor
This episode is brought to you by Indeed. When your computer breaks, you don't wait for it to magically start working again. You fix the problem. So why wait to hire the people your company desperately needs? Use Indeed's sponsored jobs to hire top talent fast. And even better, you only pay for results. There's no need to wait. Speed up your hiring with a $75 sponsor job credit@ Indeed.com podcast. Terms and conditions apply.
- [11:19]
Abercrombie Sponsor
Abercrombie's Viral Denim sale is back and Spotify listeners get an extra 15% off with code Spotify AF. Abercrombie is known for their denim with 30 to 50% off all jeans. Find out how denim should fit. Shop the viral Denim sale in the Abercrombie app online or in stores. Valid in stores and online through August 11, 2025 in US and Canada. Excludes clearance price Reflects discount code. Valid in US and Canada through August 11, 2025. Exclusions apply. See details online.
- [11:49]
Brian McCullough
Well, let's go ahead and use that as an excuse to continue our tracking of the fallout from the release of GPT5. Noted critic of some of the AI boosterism, Gary Marcus, had a piece up this weekend saying that GPT5's release was underwhelming, offering incremental improvements and failing to meet expectations, thereby showing that pure scaling simply isn't the path to AGI. GPT5 may be a moderate quantitative improvement, and it may be cheaper, but it still fails in all the same qualitative ways as its predecessors. On chess, on reasoning, in vision, even sometimes on counting and basic math. Hallucinations linger. Dozens of shots on Goal grock, Claude, Gemini, etc. Have invariably faced the same problems. Distribution Shift has never been solved. That's exactly what it means to hit a wall, and exactly the particular set of obstacles I described in my most notorious prescient paper in 2022. Real progress on some dimensions, but stuck in place on others. Ultimately, the idea that scaling alone might get us to AGI is a hypothesis. No hypothesis has ever been given more benefit of the doubt, nor more funding. After half a trillion dollars in that direction, it is obviously time to move on. The disappointing performance of GPT5 should make that enormously clear. Pure scaling simply isn't the path to AGI. It turns out that that attention, the key component in LLMs and the focus of the justly famous transformer paper is not, in fact, all you need. All I am saying is give neuro symbolic AI with explicit world models a chance. Only once we have systems that can reason about enduring representations of the world, including but not limited to abstract symbolic ones, will we have a genuine shot at AGI. He had a lot more to say than just what I quoted you there, so click through on the link to read the whole thing. And also over the weekend David Sachs of all in fame had a long tweet that a ton of people responded to, including Elon Musk, Marc Benioff and based Beff Jesos. And this is what Sachs said the doomer narratives were wrong. Predicated on a rapid takeoff to AGI, they predicted that the leading AI model would use its intelligence to self improve, leaving others in the dust and quickly achieving a godlike superintelligence. Instead we are seeing the opposite. The leading models are clustering around similar performance benchmarks. Model companies continue to leapfrog each other with their latest versions, which shouldn't be possible if one achieves rapid takeoff. Models are developing areas of competitive advantage, becoming increasingly specialized in personality modes, coding and math as opposed to one model becoming all knowing. None of this is to gainsay the progress. We are seeing strong improvement in quality, usability and price to performance across the top top model companies. This is the stuff of great engineering and should be celebrated. It's just not the stuff of apocalyptic pronouncements. Oppenheimer has left the building. The AI race is highly dynamic, so this could change. But right now the current situation is Goldilocks. We have five major American companies vigorously competing on frontier models. This brings out the best in everyone and helps America win the AI race. As Balaji has written, we have many models from many factions that have all converged on similar capabilities rather than a huge lead between the best model and the rest. So we should expect a balance of power between various human AI fusions rather than a single dominant AGI that will turn us all into paperclipspillars of salt. So far we have avoided a monopolistic outcome that vests all power and control in a single entity. In my view, the most likely dystopian outcome with AI is a marriage of corporate and state power similar to what we saw exposed in the Twitter files, where trust and safety gets weaponized into government censorship and control. At least when you have multiple strong private sector players, that gets harder. By contrast, winner take all dynamics are more likely to produce Orwellian outcomes. End quote. He went on to say that he thinks open source will play a key role going forward offering most capabilities at a fraction of the cost, which is appealing for customization and control. AI's value will split between general models and vertical applications, with humans providing goals, context and verification. He thinks that job loss fears are overblown because success favors those who use AI effectively. And again, read the whole thing because he said a lot more. So link to that one in the show notes finally today, the end of another tech era AOL is discontinuing its dial up Internet service on September 30th after launching it all the way back in 1991. Quoting PC Gamer, AOL announced this week that it will be discontinuing its dial up service next month, which probably isn't as surprising as the fact that AOL was still offering dial up in 2025 to begin with. Dial up usage has plummeted over recent years. AOL had some 1.5 million dial up users in 2015, but as of 2021 that number was reportedly in the low thousands. So this change probably doesn't impact many people. But it's somewhat shocking that dial up usage has carried on for so long. AOL originally launched its dial UP service in 1991, meaning it will be 34 years old when it's finally shut down next month. Unfortunately, some of the people still using dial up probably don't have many other options for Internet access. Broadband infrastructure has yet to make its way to some remote rural areas or if it is present, it's expensive. Many of the AOL customers who were still subscribed to its dial up service likely stuck with it out of necessity. Luckily for those people, AOL isn't the only dial up provider around. There are still a few others offering the antiquated service, however. Ideally, efforts are made to continue expanding broadband access so people don't have to keep leaning on slow, outdated Internet options like dial up. An estimated 22.3% of Americans in rural areas and 27.7% of Americans in tribal lands still lack access to high speed Internet. In urban areas. Only 1.5% of Americans face the same issue. End quote. Here is London Giddy London. Is it home of the free or what? I guess I picked a great week to come back here. The weather today was insane. Talk to you tomorrow.