A (0:04)

Welcome to the Tech Brew Ride home for Friday, October 10, 2025. I'm Brian McCullough. Today, China is stirring the pot on the tech trade wars once again. SORA has grown faster than even ChatGPT did. Is OpenAI now better at Vibe coding than Anthropic? Is how just a handful of malicious documents can poison an LLM. And in the long reads the one game that has the fate of EA on its shoulders. Here's what you missed today. In the world of tech, when your business evolves, so does your risk of data loss. But with Veeam, your data is always on the map. Partner with Veeam for coverage that keeps you moving and get protection for workloads of all shapes and sizes, even the ones you haven't created yet so you can stay resilient as you scale. With Veeam, it's all good. Get workload coverage that works for your business. @veeam.com that's V E E A M.com China is getting tetchy again. Sources tell the FT that China has tightened customs checks on chip imports, starting with Nvidia's H20 and RTX Pro 6000D after urging local tech companies to avoid Nvidia products. Quote Teams of customs officers have been mobilized at major ports across the country in the past few weeks to carry out stringent checks on semiconductor shipments, according to three people with knowledge of the matter. The inspection started with the goal of ensuring that local companies stop ordering Nvidia's China specific chips, following guidance from Chinese regulators to discourage their purchase, said the people. One person said the checks had been extended more recently to all advanced semiconductor products to also better target the smuggling of high end chips that breach US Export curbs. Chinese customs had previously done little to prevent chip imports as long as appropriate duties were paid at the border. The Financial Times reported that at least $1 billion worth of Nvidia's top AI chips were smuggled and sold in in the three months from May end quote China is also banning Tech Insights from working with or receiving data from Chinese entities, citing national security concerns. Quoting cnbc China's Commerce Ministry, citing national security concerns, announced Thursday that Tech Insights was designated an unreliable entity, which prohibits Chinese individuals or organizations from sharing information with the Canadian based company. Tech Insights is well known in the global tech space for its in depth coverage of Chinese made chips and was among the first to report breakthroughs by companies like Huawei. Beijing's crackdown on Tech Insights came less than a week after the firm revealed that a breakdown of Huawei's latest artificial intelligence chips found components sourced from outside mainland China. The findings by Tech Insights about Huawei's latest Ascend AI chips were consistent with those from other research firms like Semianalysis, which said that the Chinese company relies on technology from memory chip makers like Samsung and contract chip maker Taiwan Semiconductor Manufacturing. End quote well, maybe it's tit for tat time again, because the Senate here in the US has passed a measure requiring Nvidia and AMD to prioritize US Customers over China for advanced AI chip sales as part of a defense policy bill that is making its way through Congress. Bill Peebles, head of Sora at OpenAI, says the app hit 1 million downloads less than five days after its launch on September 30, which is even faster than ChatGPT exploded. One thing to note, though, is there has been some back and forth on the whole copyright issue. Remember how OpenAI initially said rights holders had to proactively request that their IP not show up in Sora, otherwise it would well, they've backtracked a bit on this and from my testing, the likeness free for all has been a bit curtailed because there has been some pushback. Quoting cnbc, the Motion Picture association, which advocates on behalf of the television, motion picture and home video industries, said in a statement Monday that videos that infringe our members, films, shows and characters have proliferated on OpenAI's service. OpenAI needs to take immediate and decisive action to address this issue, charles Rifkin, CEO of the mpa, said in a statement. Well established copyright law safeguards the rights of creators and applies here, end quote. OpenAI CEO Sam Altman said the company will soon give rights holders more granular control over character generation, according to a blog post last week. During a briefing with reporters at the company's Dev Day event on Monday, Altman said some users have complained that Sora is too restrictive. He asks for patience as the company irons out best practices. Please give us some grace, altman said. The rate of change will be high, end quote Data from more than 300,000 pull requests shows OpenAI is catching up to Anthropic in AI coding, at least in terms of performance. Codex apparently has a 74.3% success rate versus Claude Code's 73.7% in code approvals. Quoting the information for Anthropic, winning the coding race seems existential. Its coding technology is the driving force behind its revenue, which mostly comes from sales of its AI models through an application programming interface to customers such as Microsoft cursor and lovable. OpenAI has a ChatGPT business juggernaut that is less dependent on coding, but its leaders view coding as a key cog in developing artificial intelligence and last year doubled down on efforts to improve its models in coding. As we reported, a big reason behind the improvement in OpenAI's coding AI performance is the release last month of GPT5 codecs for coding. Before that release, OpenAI's codecs model had a 69% success rate for code it produced, said Brexton Pham, co founder of modu. Despite the improved performance that followed the latest codecs release, Codex still significantly trails claude code in usage. The percentage of merged pull requests created with codecs through modu is 24.9%, compared to 32.1% for Claude code. That's an improvement of 5 percentage points compared to before the release of GPT5 codecs, Pham said. By another metric, claude code is even further ahead of codecs when it comes to Usage. According to NPM, a software registry, Claude Code currently has more than 5 million weekly downloads, compared to 190,000 for codecs. Codex has gotten better at coming up with a game plan for more complex coding tasks, and it's cheaper than claude code based on the data Modoo collects from its customers. Many of Modu's customers access OpenAI and Anthropic Smart models for coding through their APIs, he said. Interestingly, though, Pham pointed out that cost isn't as big of a factor as you might think for developers when choosing their coding assistant. Developers are willing to pay more right now because many of them believe costs will drop over time, though some initial data hasn't quite shown that to be the case. And for CEOs whose companies use such products, it's a lot cheaper to pay for coding assistants to boost existing software engineers than it is to hire more human software engineers, Fam said. Modu's data show other surprising takeaways. The coding assistant, whose output is most likely to be accepted by developers with a 76.8% acceptance rate, is SourceGraph's AMP agent. The San Francisco based startup isn't an underdog by any means. It's raised more than $220 million in funding and was last valued at $2.6 billion in 2021. But it's definitely less hyped than the other coding agents. End quote Speaking of In a memo to staff, Vishal Shah, Meta's VP of Metaverse, has told his team to use AI to go 5x faster and expects 80% of staff to integrate AI into their day to day by Q4, quoting Wired the idea is that programmers should be using AI to work five times more efficiently than they are currently working, not just using it to go 5% more efficiently. Our goal is simple, yet audacious. Make AI a habit, not a novelty. This means prioritizing training and adoption for everyone so that using AI becomes second nature just like any other tool. We rely on the message read. It also means integrating AI into every major code base and workflow. Shah added that this doesn't just apply to engineers. I want to see PM's designers and cross functional partners rolling up their sleeves and building prototypes, fixing bugs and pushing the boundaries of what's possible, he wrote. I want to see us go 5x faster, faster by eliminating the frictions that slow us down and 5x faster to get to how our products feel much more quickly. Imagine a world where anyone can rapidly prototype an idea, and feedback loops are measured in hours, not weeks. That's the future we're building, end quote. In his message, Shah said that we expect 80% of Metaverse employees to have integrated AI into their daily work routines by the end of this year, with rapid growth in engineering usage and a relentless focus on learning from the time and output we gain. He went on to reference a series of upcoming trainings and internal documents about AI coding, including two Metaverse Day of AI learning events. Dedicate the time, take the training seriously, share what you learn, and don't be afraid to experiment, he added. This is kind of shocking. A study has found that as few as 250 malicious documents can produce a backdoor vulnerability in an LLM, regardless of model size or training data volume. Quoting the register Poisoning AI models might be way easier than previously thought if an Anthropic study is anything to go on. For those unfamiliar with AI poisoning, it's an attack that relies on introducing malicious information into AI training data sets that convinces them to return, say, faulty code snippets or exfiltrate sensitive data. The common assumption about poisoning attacks, Anthropic noted, was that an attacker had to control a certain percentage of model training data in order to make a poisoning attack successful. But their trials show that's not the case in the slightest, at least for one particular kind of attack. Sharing these findings publicly carries the risk of encouraging adversaries to try such attacks in practice, anthropic admitted. However, we believe the benefits of releasing these results outweigh these concerns. Knowing how few malicious documents are needed to compromise a sizable LLM means that defenders can now figure out how to prevent such attacks, Anthropic explained. The researchers didn't have much to offer in the way of recommendations, since that wasn't in the scope of their research. Though they did note that post training may reduce the risk of poisoning, as would continued clean training and adding defenses to different stages of the training pipeline, like data filtering and backdoor detection and elicitation. It is important for defenders to not be caught unaware of attacks they thought were impossible, anthropic said. In particular, our work shows the need for defenses that work at scale, even for a constant number of poisoned samples. End quote Another sign of the hotness in prediction marketplaces. Kalshi raised a $300 million Series D round at a $5 billion valuation as it expands to 140 countries and is on track for $50 billion in annualized trading volume this year, up from 300 million in 2024. That's significant, quoting the Times. The fundraising announcement in the same week that polymarket, Kalshi's chief rival, said that the parent of the New York Stock Exchange would invest up to $2 billion underscores how big a business prediction markets have become. Last month, Kalshi overtook Polymarket to become the biggest player in the prediction market industry, with a global market share, a prediction market activity of more than 60%, according to the data provider Dune. We have not expected this level of growth, tarek Mansour, a Kalshi co founder and its chief executive, said in an interview. That rapid expansion drew the attention of venture capital firms including Sequoia Capital, which first backed the company in 2020, and Andreessen Horowitz, a new investor. Despite closing its previous investment round in June, Kalshi negotiated the new fundraising effort, which more than doubled its valuation in August. Kalshi's recent growth has been powered by sports, especially after it began offer kind of complex wagers known as parlays. Its push into sports has rattled the traditional sports books, with the shares of DraftKings and the parent of FanDuel falling by double digit percentages over the past month. Kalshi has also benefited from a series of deals that allow customers of brokerages like Robinhood and Webull to bet on contracts directly from those sites as easily as they can buy stocks. Until now, Kalshi has operated only in the United States. Polymarket, on the other hand, was barred from letting Americans participate on its platform in 2022 and only recently resumed doing business in the country. Donald Trump Jr. President Trump's eldest son, is an advisor to both companies, joining Kalshi in January and polymarket in August. In May, the US Commodity Futures Trading Commission dropped a legal challenge to Kalshi's election related betting business, but the company is now facing pushback from US State regulators over sports betting. Several states have sued the company, arguing that it is essentially skirting their rules around online sports gambling, which remains illegal in 20 states, by offering financial products tied to the outcome of sporting events. Every time there's a new type of financial innovation, there's always a series of questions around regulation, Mr. Mansour said. If there weren't questions, he added, then Quote what you're doing is probably not meaningful or innovative enough. End quote Only one long read for you this weekend. It's from the Verge. It's a look at EA's Battlefield 6, coming October 10th. The launch of the game was already set to be a pivotal moment for the series and for EA itself, but especially so now as EA is set to go private in that $55 billion deal. Quote EA's fate is tightly tied to battlefield, New York University Games professor Joost van Drunen tells The Verge. If Battlefield 6 underperforms, it will call into question the strength of EA's non sports portfolio. At that point, the company faces pressure to rethink its strategy, perhaps even to divest and lean more heavily on the dependable sports business. It makes Battlefield a bellwether that will tell us if EA can still compete in blockbuster shooters. The pressure for Battlefield 6 started building almost immediately after Battlefield 2042's disastrous 2021 debut. The multiplayer only 2042 added huge 128 player battles to a franchise already known for epic multiplayer skirmishes. But the game had a rough, bug filled launch that necessitated multiple patches shortly after release. EA moved quickly to put Vince Zampella, co founder of Apex Legends maker Respawn Entertainment and former head of Call Duty studio Infinity Ward, in charge of the Battlefield series to help right the ship. But the damage was already done. A few months after 2042 came out, EA CEO Andrew Wilson put things bluntly on an earnings call. The launch of Battlefield 2042 did not meet expectations, even though EA is also known for titles like Mass Effect and Plants versus Zombies. Most of its money comes from a small number of franchises, along with add ons like microtransactions and battle passes. Live service revenue is already big money for EA. According to its most recent quarterly earnings report, 83% of its net revenue came from live services and other if Battlefield 6 is successful over the long term that potentially makes the game a big cash cow for ea. No weekend bonus episode this weekend, but Monday is a holiday here in the US So there will be no regular show on Monday, but I will have a bonus episode for you that day. An interview with the CEO of Rent the Runway. We don't get to talk about fashion much on this show, so some fashion tech talk to you on Tuesday.