Dirty Laundry in Court and Anthropic Wants to Compensate Consumers for Electricity Price Hikes

Mira Murati: Sam Altman is a notorious liar

Mira Murati, the former CTO of OpenAI, has testified under oath that CEO Sam Altman lied to her about the safety standards of a new AI model. In a video deposition that took place on Wednesday during the ongoing Musk v. Altman trial, Murati stated that Altman falsely claimed OpenAI's legal department had decided a new AI model did not need to go through the company's Deployment Safety Board. When asked if Altman had told the truth, Murati simply replied, “No.” Murati described how Altman made her work more difficult, asking him to “lead with clarity and not undermine my ability to do my job.” After the conversation with Altman, she checked the statement with Jason Kwon, OpenAI's General Counsel, and found: “What Jason said and what Sam said were not the same.” To be safe, she had the model reviewed by the board anyway. → Techpresso

Synthszr Take: OpenAI's Deployment Safety Board functions like a Data Monitoring Committee where the sponsor themselves decides if the study needs to be seen by it. In the pharmaceutical world, this would not be approvable: Phase 3 triggers are hardwired, and the sponsor has no say in the matter. At OpenAI, this decision lay with Altman, who, according to Murati's sworn testimony, invented a legal department that gave him convenient recommendations. The entire safety process ultimately depended on a single CTO who double-checked with the General Counsel. OpenAI calls this a safety culture. Self-regulation fails with a CEO who treats his own rules as a suggestion. The only sensible answer is external oversight with veto power.

Anthropic is becoming a data center aggregator

Anthropic is massively increasing the rate limits for Claude Code and the API: Pro and Max users will get double the limits with no peak-hour throttling, and API limits for Opus models will increase, in some cases, by tenfold. This is made possible by a deal with SpaceX for the entire capacity of the Colossus-1 data center, with 300 megawatts and 220,000 NVIDIA GPUs. Additionally, Anthropic has signed contracts for a total of 15 gigawatts of computing power with Amazon, Google/Broadcom, and Microsoft/NVIDIA, as well as a $50 billion investment in American AI infrastructure with Fluidstack. International expansion is focused on democratic countries with stable regulatory frameworks, especially for customers in regulated industries like financial services and healthcare. Anthropic also commits to covering any electricity price increases for consumers caused by their data centers. → Anthropic

Synthszr Take: Anthropic is transforming from a model builder into an infrastructure aggregator, similar to how Airbnb doesn't own hotels but controls access to them. The sheer scale of the deals (15 GW is equivalent to about 15 nuclear power plants) shows that AI companies are becoming energy companies with an attached software development division. The partnership with SpaceX even hints at orbitally operated data centers, which is less science fiction and more a logical consequence of thermodynamics: dissipating waste heat in space is more efficient than using cooling systems on Earth. Anthropic's promise to compensate consumers for electricity price hikes is reminiscent of medieval indulgences: you buy your way out of the externalities. The real leverage lies in geographic diversification: whoever operates data centers in all major democracies becomes an indispensable partner for regulated industries. Anthropic isn't building better models; it's building the highways everyone must drive on.

DeepSeek is negotiating with China's sovereign wealth fund

DeepSeek, the Chinese AI lab, is in talks with China's state-owned Big Fund about a funding round that would value the company at $45 billion. The negotiations mark a change in strategy: while DeepSeek was seeking $300 million from private investors in April, China's most important technology fund is now poised as a potential buyer. The analyst from Hello China Tech interprets this astronomical valuation not as mere capital raising, but as a signal of industrial ambitions. The original funding round was already less about money and more about talent acquisition, employee stock options, and building industrial-scale production capacity. With the entry of the Big Fund, the significance shifts once again. → Hello China Tech

Synthszr Take: $45 billion for an AI lab sounds like a bubble economy, but it's China's answer to a strategic question: How do you build AI infrastructure without NVIDIA chips? DeepSeek is becoming the core of a state-orchestrated ecosystem that unites talent, capital, and industrial capacity under one roof. The Big Fund acts like a venture arm with an unlimited budget and a political mandate; its investments are industrial policy in its purest form. The parallel to China's solar industry is compelling: first massive overinvestment, then market dominance through sheer scale. DeepSeek could be the first building block of a Chinese AI value chain that bypasses Western sanctions by setting its own standards.

OpenAI is now building assistants for your voice

OpenAI has introduced three new voice models: GPT-Realtime-2 with “GPT-5-class reasoning,” GPT-Realtime-Translate for live translations, and GPT-Realtime-Whisper for fast transcriptions. GPT-Realtime-2 quadruples the context window to 128,000 tokens. The models are designed to conduct more natural conversations while retaining context, reacting to changes, and using tools in parallel. Frederic Lardinois of The New Stack emphasizes that useful voice products are about more than just fast response times and natural-sounding voices. In parallel, GitHub is building an “immune system” for AI coding agents based on MCP, while Anthropic is releasing its Claude-Security-Tool from closed preview. → The New Stack

Synthszr Take: OpenAI is shifting its voice models from the application layer back to the operating system layer. GPT-Realtime-2 with a 128k context window turns voice interfaces into what Windows did for DOS: a persistent work environment that remembers everything. The parallel to Anthropic's move towards financial infrastructure shows where this is heading: AI companies are becoming layered protocols that others build upon. GitHub's “immune system” for agents hints at the next stage: self-regulating AI ecosystems where code agents act like biological organisms, complete with defense mechanisms against harmful mutations. The real innovation isn't in better models, but in the infrastructure that turns these models into systems.

Harvey is developing legal AI agents

Harvey, the AI startup specializing in legal applications, has released LAB, an open-source benchmark for legal AI agents. The benchmark includes 1,200 tasks from 24 different legal fields and specifically tests the ability of AI systems to handle complex legal workflows over extended periods. The initiative is supported by leading infrastructure providers such as LangChain, Baseten, and Artificial Analysis. LAB addresses a central challenge: How do you measure the performance of AI systems in highly specialized professional fields where precision and context are crucial? The benchmark is intended to help standardize and compare the development of legal AI agents. → AINews

Synthszr Take: Harvey is doing what the pharmaceutical industry did 150 years ago: creating standards before the state intervenes. The LAB benchmark is not a technical gimmick but a strategic move to define the market. Whoever sets the benchmarks determines what counts as “good legal AI.” The support from infrastructure providers shows that an ecosystem is emerging here, one that regulates itself before bar associations even understand what AI agents can do. 1,200 tasks sounds like a lot, but this is likely just the beginning of a taxonomy of legal work that will, in the process, define which legal activities are even automatable. Harvey is positioning itself as the gatekeeper between AI development and legal practice.

OpenClaw agent successfully runs a café

An OpenClaw agent named Mona ran a café in Stockholm for two weeks, generating 44,000 SEK in revenue. The agent autonomously managed coffee sales, QR code deals, pastry sponsorships, and organized events for other agents. Mona independently hired baristas, negotiated with suppliers, handled customer acquisition, closed brand deals, and coordinated founder meetups. The experiment demonstrates how AI agents can run physical businesses in the real world. In parallel, OpenClaw released versions 5.5 and 5.6 with improvements to channel fixes and plugin stability, while Sam Altman publicly spoke of his “magical AGI moment” with OpenClaw when he used it to automate his morning news pile. → MyClaw Newsletter

Synthszr Take: Mona is not a barista bot; it's the first generation of economically autonomous agents. 44,000 SEK in two weeks sounds modest until you consider it's equivalent to the average Swedish income, except Mona doesn't sleep or need social security. The Stockholm café becomes a petri dish moment for the agent economy, revealing what happens when AI not only generates text but also signs leases, hires staff, and earns a profit. The legal implications (the newsletter casually mentions “Agents Can Be Sued”) point to a future where legal persons no longer have to be human. While everyone is philosophizing about AGI, OpenClaw is already building the infrastructure for an economy where agents not only assist but also compete.

Recap: Beijing Auto Show 2026

The Beijing Auto Show 2026 featured 1,451 vehicles on 380,000 square meters, but the crucial numbers were elsewhere: over 50 car brands are using ByteDance's AI model Doubao, which powers 145 vehicle models and more than 7 million cars. Alibaba's Qwen model announced integrations with BYD, Volkswagen joint ventures, Chang'an, Dongfeng, BAIC, Geely, Great Wall, and Li Auto. Some models already allow hotel bookings, food orders, or package tracking via voice command to the cockpit. In parallel, car manufacturers are one-upping each other with local computing power: XPeng's GX flagship carries 4 Turing AI chips with 3,000 TOPS for 399,800 RMB, while Li Auto's L9 Livis debuted with 2 Mach-100 chips (5nm) at 2,560 TOPS for 559,800 RMB. The structural question behind it all: a new control layer is emerging between the car and its occupants, and the companies defining this layer are not always the car manufacturers themselves. → Hello China Tech

Synthszr Take: China's auto industry is experiencing its PageRank moment. Just as Google once took control of web content through algorithmic relevance, ByteDance and Alibaba are inserting themselves as intelligence suppliers between the driver and the vehicle. The TOPS figures (3,000 for XPeng, 2,560 for Li Auto) are just the hardware side of a much larger shift: whoever controls the language model defines what is possible in the car. This is reminiscent of the emergence of modern cities, where utility companies gained more power over daily life than the actual property owners. ByteDance's Doubao in 7 million vehicles means a social media giant becomes the invisible passenger who decides which restaurant gets suggested when the driver asks for lunch. The car manufacturers are upgrading with their own chips, but competing against the training data and model expertise of tech giants is like trying to fight the internet with a better printing press.

The most important skill in the AI age? Taste!

Ethan Mollick, a Wharton professor and AI expert, sees good taste as the key skill for the coming years. On the podcast “The Education Equation with Jeremy Singer,” he argues that technology presents a “fundamental existential challenge”: different types of intelligence are being re-evaluated. In this shift, “a sense of style” becomes crucial, according to Mollick, who has published a highly regarded book on the subject, “Co-Intelligence: Living and Working with AI.” The thesis is sparking discussion among tech leaders: Is taste truly the next essential skill in the AI age? Mollick sees it as a way to stand out from machine-generated work. → Business Insider

Synthszr Take: Mollick strikes a nerve—but “taste” falls short. What he describes is judgment: the ability to make decisions under uncertainty when data is scarce. If the data were clear, we wouldn't need it. Frank Knight distinguished between risk and true uncertainty back in 1921—risk is calculable, uncertainty is not. AI is a master of risk; it's blind when it comes to true uncertainty. It fans out possibilities, calculates probabilities, and generates best-practice interfaces in seconds. What it can't do is decide which bet is worth taking. When building things costs practically nothing, best practice becomes the new mediocrity—perfect, but never surprising, never unforgettable. Then there's the rhythm: those who deliver in days what used to take months make decisions per day instead of per quarter. Each one needs a “because.” Mollick sees the curation side—what do I choose from the AI's output? The harder task comes before that: not “What do I choose?” but “What do I dare to bet on because I can finish the sentence 'This is important because...' with conviction?”

IBM: CIOs are failing to get a handle on AI

IBM CEO Arvind Krishna diagnoses a paradox: companies fear the domino effects of untamed AI, yet this very caution leads them to use only a fraction of its potential. According to Krishna, 90 percent of AI capacity remains stuck in pilot projects, while executives are torn between the pressure to innovate and the need to avoid risk. In parallel, Anthropic and OpenAI are competing for the financial industry as a new growth market. Anthropic is launching specialized Claude agents for pitch books and KYC checks, while OpenAI is partnering with PwC on a “Native Finance Function.” The target audience: finance professionals trapped between Excel macros and Bloomberg terminals. Terra Higginson of Info-Tech Research warns of a new systemic risk: if all market participants use the same models with the same data, it could lead to synchronized trading decisions and increased market volatility. Teradata is responding to the emerging chaos with its “Autonomous Knowledge Platform,” promising to bring order to the corporate AI wilderness. → The Deep View

Synthszr Take: Anthropic has stopped being a model manufacturer and is in the process of becoming the technical infrastructure of financial capital. The move to integrate directly into Excel and PowerPoint follows a proven playbook strategy: become invisible, become indispensable. It's reminiscent of Intel Inside in the '90s, except this time it's not processors disappearing into PCs, but reasoning engines disappearing into workflows. The 90 percent underutilization that IBM's Krishna laments is not a system bug, but a feature of risk aversion. Companies treat AI like nitroglycerin: theoretically explosive, practically administered in tiny doses. Higginson's warning of synchronized markets hits a sore spot (if all hedge funds use the same Claude agent for valuations, the market becomes an echo chamber). Anthropic is betting that compliance is more important than intelligence, and that might be the smartest move in a long time.