Weekend Update: Hy4, Big Tech Profits, and Data Center Robots

Compact Conversations for 2026-08-30: 5 AI stories, ai news worth knowing in just 5 minutes.

[Audio embed placeholder]

The Lead: Tencent Releases Hy4 Preview, a Massive Open-Weight LLM

Tencent has released a preview of Hy4, a text-only open-weight model with 770 billion total parameters, 49 billion active parameters, and a 1 million token context window. It’s a significant scale-up from last month’s Hy3 model and includes a unique ‘reasoning effort’ setting with just ‘high’ or ‘no_think’ modes.

Why it matters: This release signals continued rapid scaling in the open-weight model space from a major player, offering developers a new, extremely large model to experiment with. The explicit reasoning control could influence how future models are architected for different computational budgets.

Source: Simon Willison’s Weblog

The Feed

Big Tech Profits Get $160 Billion Boost from AI Investment Gains

Paper gains on stakes in companies like OpenAI, Anthropic, and SpaceX have added an estimated $160 billion to Big Tech’s reported profits, complicating the analysis of their core operational health.

Why it matters: These large, unrealized investment gains are distorting traditional earnings metrics, making it harder for investors and analysts to assess the underlying performance of tech giants’ primary businesses.

Source: Financial Times

Anthropic Warns of Infostealer Malware Hijacking Claude Sessions

Anthropic is notifying users that infostealer malware has stolen active session cookies to hijack Claude accounts, leading to drained usage credits. The company has responded by signing users out, removing saved payment methods, and issuing refunds.

Why it matters: This highlights a growing enterprise security concern: AI service accounts with usage-based billing are becoming targets for credential theft, requiring vigilance around endpoint security and session management.

Source: BleepingComputer

Meta Tests Robots for Data Center Tasks Like Cable Management

Meta is conducting tests using robots from vendors like Watney Robotics and ABB to perform physical tasks in data centers, including plugging in cables and resetting servers, as part of a broader industry push to automate operations.

Why it matters: Automating physical data center work could significantly impact operational costs and staffing models for hyperscalers, especially as AI infrastructure scale makes efficiency and 24/7 operation more critical.

Source: Ars Technica

oMLX: A Mac LLM Server Aims to Slash Agent Wait Times

oMLX is a new Mac application that serves as a local LLM server, claiming to reduce agent wait times from 90 seconds to 5 seconds by managing multiple open-source models as a single, fast endpoint.

Why it matters: For developers building or testing with local models, a fast, unified local server can streamline workflows, reduce latency, and lower costs compared to frequent cloud API calls.

Source: Product Hunt

One Thing to Try

Instead of asking an AI ‘is this code correct?’, try this two-step prompt sequence to force a context switch and uncover hidden bugs. First, ask the model to explain the code’s function, assumptions, and silent failures as if to a new developer. Then, prompt it to role-play as a skeptical reviewer tasked with listing every possible failure mode.

Sources

Transcript

Host A: Welcome to Compact Conversations, the show that compresses the day’s AI news into 5 minutes.

Host A: [curious] For the weekend update, Chinese tech giant Tencent released a preview of Hy4, a new open-weight language model. It’s text-only, with 770 billion total parameters and 49 billion active on any given task. It also has a one million token context window, and the full model is a 1.56 terabyte download on Hugging Face.

Host B: That’s a significant jump from their previous model, Hy3, which came out just last month. Hy3 had 295 billion total parameters and a 256,000 token context. The new Hy4 preview includes an interesting feature in its chat template: a ‘reasoning effort’ setting that only has two modes, ‘high’ or ‘no_think’. Simon Willison tested it with a creative prompt to generate an SVG of a pelican riding a bicycle, and the model’s internal reasoning trace showed it debating whether to add a helmet or sunglasses before deciding against it.

Host A: One number to know today is 160 billion dollars. That’s the amount, according to a Financial Times report, that paper gains on investments in AI companies like OpenAI, Anthropic, and SpaceX have added to Big Tech’s reported profits. [with emphasis] Analysts say these large, unrealized windfalls are muddying the tech sector’s earnings metrics. For some companies, these investment gains now represent a substantial portion of their overall reported profit, which makes it harder to assess the true health of their core operations.

Host B: In security news, Anthropic is warning some Claude users that infostealer malware on their personal computers has been hijacking active login sessions. According to BleepingComputer, the company has signed affected users out globally, removed saved payment methods, and issued refunds for usage drained by the malware. The report notes the attackers are using stolen session cookies to gain access, a common tactic for info-stealers like RedLine and Vidar.

Host A: [lighter] Next, a quick tool note. A new Mac LLM server called oMLX launched on Product Hunt. Its claim is that it can cut agent wait times from 90 seconds down to 5 seconds. The developers say it’s designed to run multiple open-source models locally—like Llama or Mistral—and manage them as a single, fast endpoint. This could be useful for testing different models in development workflows without constant cloud calls and latency.

Host B: And finally, WIRED reports that Meta is testing robots to perform physical tasks inside its data centers, like plugging in cables and resetting servers. The company is using hardware from vendors including Watney Robotics and ABB. One worker estimated a successful bot could replace up to 80 percent of some technician workloads. [thoughtful] Meta’s official statement, though, emphasizes a need for more skilled workers, not fewer, to manage the growing scale of AI infrastructure. The report says these tests are happening at sites in Iowa and Ohio, and are part of a broader, quiet industry push to automate data center operations, which are becoming a major cost center.

Host A: [conversational] One thing to try if you use AI to write code: a Reddit user shared a five-prompt sequence they run on every chunk of AI-generated code before trusting it. The core idea is that asking ‘is this correct?’ in the same chat is useless because the model will just defend its own work.

Host B: The first two steps are the most practical. First, paste the code and ask the model to explain it back to you as if you didn’t write the request—highlighting what it assumes about inputs and what it silently doesn’t handle. Then, in a new message, tell the model to act as a reviewer who believes there’s a bug and must list every possible way the code could fail: bad inputs, empty cases, concurrency issues. Rank them by likelihood. The user says this two-step reset forces a context switch and often surfaces edge cases or logic flaws that a simple verification prompt would miss.

Host A: That’s Compact Conversations for Sunday. More AI news tomorrow. Until then, happy prompting.