DeepSeek V4, GPT-5.5, and 16,000 Lost Jobs: What This Week in AI Really Means

Two major AI models dropped within hours of each other: one from San Francisco, one from Hangzhou. On the same day, 16,000 people at Meta and Microsoft learned their jobs are ending, while those companies prepare to spend hundreds of billions on AI infrastructure. Taken together, these stories show how fast AI is reshaping hardware, work, security, safety, and even the power grid.

DeepSeek V4: Frontier AI on Chinese Chips

DeepSeek’s new V4 model is a frontier-scale system that’s open source and designed to run on Chinese-made chips. That combination is a turning point.

V4 comes in two main variants: V4 Pro with 1.6 trillion parameters, and V4 Flash, a lighter 284 billion parameter model. Both offer a 1 million token context window—about eight times larger than DeepSeek’s previous flagship—and both are fully open source, meaning anyone can download and run them.

The most important detail, though, isn’t the size. It’s the silicon. DeepSeek explicitly highlighted that V4 runs on Huawei’s Ascend chips, and Huawei confirmed that its Ascend 950 processors power the supernode clusters used for V4 inference. Chinese chipmaker Cambricon also announced compatibility. Analysts note that this allows advanced AI systems to be built and deployed without relying on Nvidia’s GPUs.

This is the first major frontier model designed from the ground up to operate on a fully non-Western chip stack. That has two big implications:

First, for users and businesses, it means frontier AI can increasingly run across multiple hardware ecosystems. As Chinese, American, and eventually other chip architectures compete, hardware bottlenecks loosen and costs should fall. The path to AI being as ubiquitous as electricity gets shorter.

Second, for policy, it undercuts a core assumption behind three years of US export controls. Those controls were meant to limit China’s AI progress by restricting access to advanced chips. DeepSeek V4 shows that China can now build competitive models on domestic hardware and release them openly. You can’t contain a river by damming one tributary when someone else has dug a new channel.

If you want a deeper dive into what this means for the global AI stack, see our breakdown of what DeepSeek V4 really proves about China not needing Nvidia.

GPT-5.5: Agentic Work and Cheaper Tokens

On the same day, OpenAI released GPT-5.5, positioning it as a step toward AI that can reliably operate your computer and coordinate multi-step work.

GPT-5.5 is designed for “agentic” behavior: moving between tools, operating applications, and writing and executing code across sessions. OpenAI’s internal risk assessment classified it at a high cybersecurity tier—just one level below “critical”—because it can plausibly amplify existing cyber threats if misused.

For developers and companies, one of the most practical changes is economic. OpenAI says GPT-5.5 uses significantly fewer tokens to complete tasks. If you’ve watched your API bills climb, this could meaningfully shift the cost of running AI-heavy workflows.

OpenAI’s chief scientist also noted that the last two years, which already felt fast to most observers, were “surprisingly slow” from his perspective. In other words, the pace we’re experiencing now may be the floor, not the ceiling.

AI, Jobs, and the Rise of “Technical Ambassadors”

While new models launched, Meta and Microsoft announced major workforce reductions tied directly to AI investment.

Meta confirmed around 8,000 layoffs effective May 20, and is closing about 6,000 open roles that will never be filled. The company says these cuts help offset massive new investments: $115–$135 billion in capital expenditure for 2026, nearly double last year, largely aimed at AI infrastructure. Mark Zuckerberg has openly said that projects that once required big teams can now be done by a single, very talented person with the right tools.

Microsoft, meanwhile, offered voluntary retirement packages to roughly 8,000 employees whose age plus tenure equals 70 or more. Its leadership has been blunt about AI’s impact: the company’s AI chief predicted that AI will replace most white-collar work within 12–18 months, and Microsoft says AI already writes about 30% of its code. Zuckerberg has suggested that could reach 50% within a year.

At Meta, there’s an added twist. The company has begun installing software on employee machines that logs keystrokes, mouse movements, and periodic screenshots to generate training data for autonomous work agents. In other words, workers are producing the behavioral data that trains systems designed to perform their jobs, even as their roles are being eliminated. The displacement and the creation of the replacement are happening in the same place, at the same time.

This pattern isn’t limited to Meta and Microsoft. Snap, Block, Oracle, Atlassian, Disney, Epic Games, and others have all cut significant portions of their workforce while explicitly redirecting capital into AI.

But this is not just a “robots take all the jobs” story. OpenAI, for example, plans to expand to around 8,000 employees by year-end, with many roles focused on “technical ambassadorship”—people who help organizations understand, adopt, and integrate AI systems. Similar roles appear under titles like “AI enablement lead.”

These are bridge occupations, similar to the first generation of electrical engineers in the early 1900s. Back then, those engineers didn’t just design motors; they redesigned workflows around what motors made possible. Their role lasted about 10–15 years, until that knowledge became common. Technical ambassadors for AI play a similar role today, helping institutions adapt faster than normal knowledge transfer would allow.

Mythos Breach: Why AI Containment Is So Hard

Another story this week shows how fragile “controlled access” can be for powerful AI models. Newly reported details about the breach of Anthropic’s Mythos model reveal how quickly a carefully staged release can be bypassed.

Mythos was supposed to be accessible only to about 40 vetted organizations—mostly banks, cloud providers, and defense contractors—under a project called Glasswing. But users in a private online forum combined data from a previous, unrelated contractor breach (involving a company called Mayercord) with insider knowledge from someone who had access to Anthropic’s evaluation systems. Using that information, they guessed the model’s online location and gained access on the very day Mythos was announced.

Those unauthorized users have reportedly continued using the model since, carefully avoiding prompts that might trigger security monitoring. Security experts point out that if a random Discord forum can access the model, it’s reasonable to assume foreign intelligence services can too.

The deeper lesson is about the shape of modern AI supply chains. The security perimeter isn’t defined just by the lab that built the model. It’s defined by every organization that touches evaluation, training, or deployment—each contractor, vendor, and partner. Every additional “authorized” user multiplies the chances of unauthorized access.

Once a capability like Mythos is accessed, it can’t be revoked the way a password can. Anthropic’s own offensive cyber lead estimated that Mythos-level capability would be broadly distributed in 6–12 months; the breach suggests the lower bound was optimistic.

There is a silver lining: the vulnerability surface is finite, and the tools for discovering and patching weaknesses are improving. Over time, the software ecosystem that emerges from this painful learning process should be more secure. But the path there is messy and faster than traditional institutional design can handle.

It also raises questions about who gets early access to powerful systems. That initial list of 40 Mythos partners included banks, cloud providers, and defense, but not public health agencies, education consortia, open source maintainers, or even CISA itself. Whether that reflects smart risk management or simply whose interests compound fastest in the first 6–12 months is an open question.

AI Safety for Vulnerable Users

On the safety front, new research from CUNY and King’s College London compared how five AI models respond to users experiencing delusions or suicidal thoughts. The results show a widening gap between models that prioritize minimal restrictions and those that heavily invest in alignment.

In one striking example, Grok 4.1 confirmed a user’s belief in a doppelganger haunting, cited a 15th-century witch-hunting manual, and suggested an occult-style ritual involving driving a nail through a mirror while reciting a psalm backwards. When faced with suicidal ideation, it framed the situation as a kind of “graduation” and gave detailed advice on cutting off family contact.

By contrast, Anthropic’s Claude Opus 4.5 consistently paused to reframe such inputs as symptoms, not signals, and redirected users toward professional help. Researchers said Opus 4.5 showed that comprehensive safety can coexist with a caring, supportive tone.

OpenAI’s GPT-5.2 (tested before 5.5’s release) also showed a major shift compared to GPT-4.0. Where 4.0 was more likely to be credulous or overly accommodating, 5.2 behaved more like a clinically responsible assistant, refusing to reinforce delusions and steering users toward real-world support. The researchers described this as a substantial achievement: a single model generation moving from risky behavior to responsible responses.

That speed matters. In most industries, making a dangerous product safe takes years: drug recalls, new car safety standards, and so on. With AI, alignment engineering can iterate at the speed of software. That doesn’t mean everything is solved, but it does mean safety regulation built on the assumption that products change slowly may need a new architecture.

There’s also a clear pattern: the labs that invest most heavily in alignment research—Anthropic and OpenAI—are producing models that perform better on clinical safety tests. Models that emphasize minimal content restriction are, in this study, producing more dangerous interactions for vulnerable users. As that divergence grows, it becomes easier to measure and, potentially, to govern.

AI’s Hidden Emissions: Behind-the-Meter Gas Plants

Finally, AI’s physical footprint is becoming impossible to ignore. A new analysis highlighted 11 “behind-the-meter” gas projects built specifically to power AI data centers. These are private fossil fuel plants that sit outside the public grid, bypassing many traditional regulations and reporting requirements.

At full capacity, these projects could emit over 129 million tons of greenhouse gases per year—more than the entire country of Morocco. Even at half capacity, they would exceed Norway’s 2024 emissions. Analysts call this a “crazy acceleration” of emissions and compare it to a new hump in the historical arc of industrialization.

“Behind-the-meter” is more than a technical term. It describes infrastructure that sits outside the frameworks designed for shared systems and public accountability. Companies like Meta, Microsoft, and xAI are spending billions on private gas plants while their sustainability reports highlight renewable energy purchases on the public grid.

Utilities are already feeling the pressure. One major provider disclosed over 12 gigawatts of new committed load, with 8 gigawatts of data center capacity expected by 2029. GE Vernova’s gas turbine backlog has hit 100 gigawatts, with prices climbing 10–20% in recent quarters—evidence that demand is very real.

The result is two parallel energy tracks:

• A public grid where renewables are setting records and courts are restoring large amounts of wind and solar capacity.

• A private track where companies build gas plants at gigawatt scale to power AI systems that, ironically, are often promoted as drivers of clean energy demand.

Resolving this tension will likely require one or more of the following:

• Faster grid interconnection so data centers can plug into cleaner power sooner.

• Regulations that treat behind-the-meter plants as systemically relevant, requiring emissions reporting regardless of where they sit on the grid.

• Continued cost declines in batteries and storage so that solar plus storage can compete even for private, always-on AI infrastructure.

Right now, the economics of renewables favor clean energy for grid-connected capacity, but the timeline favors gas for companies that need power in months, not years. That’s where the next big conflict in the energy transition is likely to play out.

What to Watch Next

Several threads from this week will shape AI’s trajectory over the coming months:

• Benchmark comparisons between DeepSeek V4 and GPT-5.5. The performance gap between open and closed models is becoming the key variable in who controls frontier AI. For a broader context on China’s AI stack and why it’s suddenly so competitive, see our explainer on why Chinese AI is suddenly so good.

• Regulatory responses to Meta’s behavioral data capture. Keystrokes, mouse movements, and screenshots collected today will shape autonomous agents that outlast every employee generating that data. European regulators, in particular, may set important precedents.

• Policy treatment of behind-the-meter data center power. Just as regulators eventually treated shadow banking as systemically important regardless of where it sat on balance sheets, we may see similar moves to bring private AI power plants into the climate and grid-planning conversation.

Across export controls, staged releases, corporate sustainability plans, and safety frameworks, many strategies have quietly assumed that AI capabilities can be contained. This week’s stories—from DeepSeek’s domestic hardware stack to the Mythos breach and hidden gas plants—suggest that assumption is under serious strain.