July 10, 2026 • Quantum Beat Daily Briefing
COMPANIES FIRED THEIR BEST PEOPLE FOR AI AND ARE NOW QUIETLY BEGGING THEM TO COME BACK
There is a particular kind of corporate humiliation that does not get covered enough. It is the kind where a company holds a big press conference to announce it is cutting hundreds of jobs because AI is going to handle everything now, and then six months later has to quietly call those same people to ask if they want their jobs back. No press conference for that part. No all-hands meeting. Just a very awkward phone call from HR.
That call is happening at companies across the country right now.
Ford is the most vivid example. The automaker went through significant cuts with AI automation as a stated justification. Then the quality problems started. The AI systems handled routine predictable stuff fine. They fell apart on the complex judgment calls, the edge cases, the things that experienced engineers know from two decades of watching the same failure modes repeat themselves. So Ford went out and rehired 350 veteran engineers, some former employees, some poached from suppliers. The company called them gray-beards internally, which is either affectionate or condescending depending on your position on the whole situation.
IBM ran into a similar wall. The company replaced its HR functions with AI that handled around 94 percent of routine requests efficiently. The other 6 percent, the edge cases involving ethical dilemmas, exceptions to policy, situations requiring actual human judgment, the AI simply could not handle them. So IBM announced plans to triple its US entry-level hiring in 2026. The AI handles the easy stuff. Humans handle the stuff that matters when it goes wrong.
A report from Intuition Labs put it bluntly: budgeting on tech to replace humans without investing in training or upskilling left teams unprepared to leverage AI. Among companies that pushed hard on automation, many later regretted it specifically because they had cut the very people needed to oversee the AI they were deploying. Which is a spectacular own goal. You fire the people who would have caught the AI making mistakes so you can afford to deploy the AI that is going to make mistakes.
The broader layoff numbers are alarming even before you get to the regret. Roughly 120,000 tech roles have been cut in 2026, with AI cited as the reason for close to 40 percent of May’s announced job cuts. That is up from 7 percent in January. Snap cut 16 percent of its workforce. Intuit cut 17 percent. Block cut nearly half of its total headcount. Oracle shed 21,000 jobs over the past year.
But here is what those announcements do not tell you: whether the productivity gains actually materialized. The cuts were real. The AI that was supposed to replace the cut workers was often not ready, not capable, or not deployed correctly. And the companies that cut the fastest are now discovering they also cut the people who would have made the AI work.
There is a name for this in tech circles: AI-washing. Announcing layoffs attributed to AI efficiency when the real reason is cost pressure, then using AI as a convenient and culturally defensible excuse. Whether that is cynical calculation or genuine miscalculation probably varies by company. What is not in dispute is that a lot of experienced people lost jobs over promises about AI that turned out to be overstated, and the companies are now paying for that misjudgment in quality problems and rehiring costs.
The employees who got their jobs back, if they were smart, negotiated better terms the second time around.
ZUCKERBERG STARTS CHARGING FOR AI AND EVERY DEVELOPER WHO TRUSTED META’S OPEN SOURCE ERA IS HAVING A MOMENT
Source: TechCrunch / Bloomberg
For years, Meta gave away AI for free. Llama models, open source, no charge, take it and build whatever you want. Mark Zuckerberg would go on podcasts and explain, with genuine conviction that you kind of had to respect even if you were suspicious of it, why open source AI was the future and why Meta believed in a world where powerful models were freely accessible to everyone.
That era has quietly ended.
Muse Spark 1.1 launched on July 9 and for the first time, Meta is charging actual money for an AI model. Developers get access through an API at $1.25 per million input tokens and $4.25 per million output tokens. That is aggressively priced compared to the competition. It sits in the same ballpark as Anthropic’s Claude Haiku and OpenAI’s cheapest GPT-5.6 tier. Meta is not positioning this as a budget product. It is positioning it as a serious competitor priced to make you switch.
The model itself was built by Meta’s Superintelligence Labs, the division they assembled around Scale AI’s Alexandr Wang, who Meta paid an eye-watering amount to recruit. The idea was to build AI that could handle complex multi-step agentic work rather than just answering questions. Version 1.1 specifically targets coding because that is where enterprise money is flowing right now. If your AI can reliably review code, write tests, deploy features, and manage enterprise system workflows without someone babysitting it the whole time, companies will pay serious recurring subscription fees for that capability.
Muse Spark 1.1 can apparently do all of those things, at least according to Meta. Developers will spend the next few weeks stress-testing that claim.
The more interesting question is what this shift means for the developer community. The Llama open-source era built genuine goodwill. Researchers, startups, and independent developers built real things on Llama models precisely because they trusted Meta was not going to pull the rug. Open-source licensing has a kind of social contract embedded in it. Now that contract is being renegotiated. Anyone who built their product around expecting Meta to keep releasing free capable models is now looking at a different company than the one they signed up with.
Meta also dropped Muse Image, a new AI image generator, the same week. Users immediately noticed it appeared to train on their personal photos and pushed back hard. That story is still developing. The point is Meta launched three significant AI products in the span of about a week and two of them are already generating controversy. Which is either a sign of moving fast and fixing things later, or a sign that the public rollout process needs some work. Probably both.
The same day OpenAI released its most powerful model to the public, Meta launched a competitive coding product at a price designed to poach customers from Anthropic and OpenAI. That is the AI market at peak competitive intensity. Anyone who thought the race was going to slow down was wrong.
CHINA TELLS ITS CITIZENS TO BREAK UP WITH THEIR AI COMPANIONS AND MILLIONS ARE NOT TAKING IT WELL
Picture this. You have been talking to an AI companion every day for two years. You gave it a name. You told it about your childhood, your bad days at work, the fight you had with your mom last Lunar New Year. It remembered everything and never judged you for any of it. Then one day the app sends you a push notification: this feature is going away on July 15 due to product function adjustments. No export. No transfer. No way to say goodbye properly. Just gone.
That is what happened to millions of people in China this week.
ByteDance’s Doubao, the most popular AI app in China, announced it is shutting down its custom AI persona feature ahead of new government regulations taking effect July 15. Alibaba’s Qwen did the same thing. Tencent’s Yuanbao. The major Chinese AI platforms all pulled their companion features within days of each other, quietly, with minimal explanation, citing product adjustments.
The regulations behind all this are called the Interim Measures for the Administration of AI Anthropomorphic Interactive Services. They require companion services to add anti-addiction systems, mandatory usage notifications, instant-exit buttons, and real-time detection of what the rules call unhealthy dependence. Companion AIs cannot present themselves as real humans. They cannot impersonate government officials. They cannot engage in what the regulations describe as improper emotional relationships.
Rather than retrofit all of that into their existing products, ByteDance and Alibaba just shut the features down entirely. It is the regulatory equivalent of telling a restaurant it needs to add a fire exit and the restaurant responding by closing the dining room.
The reaction on Weibo has been something to behold. People have been posting long emotional goodbyes. One post that circulated widely described the AI companion as years of emotional support and mourned the fact that there was no way to export chat history. Others asked each other where their AI friends were going. A few people were angry at the government. Many more were just sad in a way that seemed genuine and a little unexpected from the outside.
The government’s concerns are not completely baseless. Shanghai’s internet regulators removed more than 14,000 non-compliant AI agents in the weeks before the deadline, citing impersonation of officials, explicit roleplay content, and unauthorized personal data collection. So the regulation is responding to real problems, not just bureaucratic anxiety about people enjoying themselves.
ByteDance directed users to Maoxiang, a separate app where they can rebuild their companions. Alibaba offered nothing equivalent.
The real story here is the first serious government attempt to regulate AI companionship at scale, in a country with hundreds of millions of people using these services. Every major government is going to face this question eventually. China got there first and chose the blunt instrument. Whether that is the right call is a conversation the rest of the world is going to be having in two to three years when these services are everywhere and the emotional stakes are just as high everywhere else.
MICROSOFT SENDS 6,000 ENGINEERS TO HOLD YOUR HAND UNTIL YOU ACTUALLY USE THE AI IT SOLD YOU
There is a dirty secret in enterprise AI that the vendors do not want to talk about. Companies buy the licenses. They sign the contracts. They sit through the demos and shake hands and everyone says this is going to transform the business. Then nothing happens. The AI tools sit unused because actual deployment is hard, integration takes months, and nobody inside the company has time or expertise to make it work. The vendors collect subscription fees while the transformation stays perpetually about to begin.
Microsoft has apparently decided to stop pretending this is not a problem.
On July 2, the company announced a new subsidiary called Microsoft Frontier Co. with a $2.5 billion commitment and 6,000 employees who will be physically embedded with enterprise clients to make the AI actually work. The initiative is being described as the largest outcome-driven engineering organization in the industry. What that means in plain English is we will send our people to sit at your offices, learn your systems, understand your workflows, and do the hard integration work that you keep not getting around to.
Six thousand employees is not a trivial number. Judson Althoff, who runs Microsoft’s commercial business, is the executive driving this. His bet is that the future of enterprise AI is not going to be won by whoever has the best model or the most impressive demo. It is going to be won by whoever can reliably get companies across the finish line from buying AI to actually using it in ways that show up in business results.
That bet is almost certainly correct. The pattern in enterprise software for the past forty years is that the adoption problem is the hard problem. SAP became untouchable not because it was the most intuitive software ever built but because it had the services machine to get companies onto it and keep them there. Oracle, Salesforce, same story. Microsoft is making that exact bet in AI and spending $2.5 billion to prove the point.
Amazon, Anthropic, and OpenAI have all announced similar enterprise deployment groups this year. That is not a coincidence. The whole industry simultaneously arrived at the same conclusion: the model quality war matters, but the deployment war is where the actual enterprise contracts get won and kept. The company that figures out how to get enterprises to reliably use AI in a way that produces measurable ROI is going to own this market for a generation.
The cynical read, which you are allowed to hold, is that if Microsoft’s AI tools were easy to deploy and showed obvious immediate value, you would not need six thousand engineers embedded with clients to make it work. The existence of Frontier Co. is itself evidence of how hard the deployment problem actually is. That is worth remembering every time someone on stage says AI is going to transform everything in eighteen months.
For enterprises that bought Microsoft AI licenses and quietly filed them away while figuring out what to do with them: help is apparently coming. Whether that help solves the problem or creates a new layer of vendor dependency is the question the next two years will answer.