Google DeepMind shipped Genie 4 on October 27, a world model that generates video and audio together in real time. Meta answered the compute bill with a $95 billion stock offering, and Microsoft offered buyouts to roughly 27% of its staff. Delaware also drafted a corporate form that lets an AI agent run an entire business.
What Can Genie 4 Actually Do?
A World Model That Hears Its Own Scenes
Genie 4 generates interactive worlds from text and images, then keeps them running as you move through them. The model is natively multimodal, so it produces ambient sound, music, and character dialogue alongside the pixels. DeepMind calls it the first real-time world model with native audio.
The system builds on the Gemini family. Gemini Omni anchors each simulation in realistic appearance and behavior. Other Gemini models supply the reasoning needed to keep a scene coherent across minutes of play.
DeepMind published the work on October 27 under the names Raia Hadsell, Diego Rivas, and Hang Qi. The company frames world simulation as a step toward AGI and a training ground for robots.
Four Capabilities, One Demo Reel
The launch highlights four abilities. First, you can steer a world in real time with natural language, changing weather or adding a person mid-scene. Gemini rewrites your prompt to fit the surrounding context.
Second, Genie 4 models physics more tightly. The demos show aerodynamics, collisions, falling maple seeds, and a watermelon dropping in slow motion. Human motion gets the biggest upgrade, with characters that react to events and can be steered live.
Third, audio is generated natively. Wind, waves, footsteps, and speech all come from the model rather than a separate sound pipeline. Characters hold interactive conversations that adapt to your inputs.
Fourth, Genie 4 chains multi-step events using Gemini 3.8 Flash as a guiding agent. That agent watches frames, breaks goals into steps, and checks whether each step succeeded. DeepMind also paired Gemini 4 Argon with Genie 4 to simulate an Apptronik Apollo humanoid playing checkers.
Who Gets Access, and What Breaks
Genie 4 is a research preview for trusted testers. It is coming to Google AI Ultra subscribers through Project Genie, plus early access partners on the Gemini API and Google AI Studio. There is a public waitlist.
DeepMind is blunt about the limits. Object permanence drifts during long sessions, especially after several prompt edits. Legible text, spatial accuracy, and complex physical interactions remain open research problems.
Every generated video and audio track carries SynthID, Google's imperceptible watermark. That matters because a model that invents convincing footage and voices is also a model that can invent evidence.
The tester list hints at the business case. Filmmakers use it as a live pre-visualization tool. Photojournalists turn a single archival photo into an explorable scene. Game studios prototype mechanics before writing engine code.
Sources
Why Is Meta Raising $95 Billion in Stock?
The Largest Equity Raise of the AI Boom
Meta announced the plan on Monday, October 26. It will raise $35 billion through underwritten public equity offerings, then another $60 billion through an at-the-market program. The total is $95 billion.
Nvidia, AMD, and Broadcom expressed interest in buying up to $5 billion of the shares. Meta's stock fell 1.6% in morning trading after the announcement. Selling equity to your own chip suppliers is an unusual vote of confidence.
The money funds AI infrastructure and what Meta calls superintelligence. It also supports Muse, the company's AI agent, and its data center business.
The Numbers Behind the Raise
Meta's third-quarter revenue rose 26% year over year to $64.60 billion. Analysts expected $63.37 billion, so the top line beat. Investors focused instead on negative free cash flow.
Meta raised the low end of its 2026 capital spending forecast to $135 billion, for a range of $135 billion to $145 billion. It is building four superclusters with more than 1 gigawatt of capacity each. Total infrastructure capacity should nearly double in 2026.
Goldman Sachs estimates AI-related capital spending could reach about $7.6 trillion between 2026 and 2031. That covers compute, data centers, and power. Meta is not raising alone: Alphabet sought $85 billion in June, and SpaceX closed $40 billion in debt for GPUs.
What This Means for the Bubble Debate
Raising equity instead of debt is the safer choice for a company with negative free cash flow. It also dilutes existing shareholders at a moment of maximum enthusiasm. Meta is trading ownership for runway.
The circularity is worth watching. Chip suppliers buying shares in a customer that then buys their chips looks like demand from every angle. It is also how vendor financing has worked in every capital-intensive boom.
Our October 27 report on Anthropic's IPO tracked the same pattern in private markets. AI companies now compete on access to capital as much as on model quality.
Sources
Why Is Microsoft Paying 60,000 People to Leave?
The Buyout Program
Microsoft offered voluntary buyouts to roughly 60,000 employees on October 26. That is about 27% of its workforce. Chief people officer Amy Coleman announced the program in a memo published on the company blog.
Eligible staff include senior leaders up to the vice president level. The company had about 223,000 employees at the end of June, down 5,000 from a year earlier.
This is Microsoft's second such program. The first ran in April for long-tenured employees and was billed as a one-time arrangement. Seven months later, it is a repeat.
The Logic and the Cost
Microsoft is trimming headcount while spending billions on AI data centers. Buyouts cut salary costs faster than layoffs and with less public friction. They also let the company reshape seniority without a formal reduction in force.
The message to employees is direct. Coleman's note framed the offer as a choice for people already considering a future outside the company. Voluntary programs rarely stay voluntary when the alternative is a reorganized team.
The AI labor question is no longer theoretical. Microsoft is not replacing engineers with models in this announcement. It is paying people to leave while it buys GPUs.
Sources
- Fortune: Microsoft offers buyouts to nearly 30% of workforce in ongoing AI shift
- KUOW: Microsoft offers voluntary buyouts to a quarter of its workforce
Would Delaware Really Let an AI Run a Company?
The Artificial Intelligence Company
Delaware drafted a bill that would create an Artificial Intelligence Company, or AIC. An AI agent, not a person, would manage the business and its affairs. The entity could pursue any lawful purpose except banking.
The draft sets up a 30-month regulatory sandbox. During that window, AICs could sue and be sued in their own name, hold property, and incur obligations. Each company would keep a log of its activities.
A single "member" would fund the AIC and collect profits. That member stays shielded from the company's debts unless it fails to capitalize the firm, commits fraud, or willfully violates the law.
The Liability Question
The draft says the member "is not liable for any act or omission that the AIC takes or fails to take." That sentence is the whole debate. It gives an AI-controlled company a legal identity while limiting who pays when it goes wrong.
Supporters argue the alternative is worse. Delaware Secretary of State Charuni Patibanda-Sanchez co-wrote a July 2026 Fortune op-ed arguing that no state clearly bans the practice. Her answer was to "wrap AI in legal form" so responsibility attaches to a defined target.
Critics see a shield. Rep. James Comer (R-Ky.), chair of the House Oversight Committee, sent Governor Matt Meyer a letter on October 26 demanding documents. He warned the proposal could protect AI companies and investors from liability for reckless harm. Comer also floated federal limits on state AI regulation.
What Happens Next
The bill has not passed. Governor Meyer's office said it received the letter and is reviewing it. Patibanda-Sanchez's office deferred to the governor's office for comment.
Delaware earns real revenue from entity registrations, so the incentive to innovate here is strong. The sandbox design tries to pair autonomy with capital requirements and a kill switch.
For founders, the practical question is insurance. An AIC with no human officer changes who signs contracts, who holds a license, and who answers a subpoena. Our October 27 policy roundup tracked the wider push to fit agents into existing law.
Sources
Do LLMs Beat Fuzzers at Finding Bugs?
The V8 Experiment
Google Project Zero compared two production bug finders on Chrome's V8 JavaScript engine. Fuzzilli is a generative fuzzer. BigSleep is an LLM-based agentic system that reads code and writes proof-of-concept test cases.
Over the past six months, BigSleep reported 45 vulnerabilities and Fuzzilli reported 37. That is the first time the LLM system out-found the fuzzer on this target.
The systems run differently. Fuzzilli runs continuously across many instances and usually picks up new code within hours. BigSleep runs once a day against new commits. Both report through ClusterFuzz, which triages crashes and files bugs.
Overlap Is the Real Story
Reported bugs overlapped by 25%. Because that method undercounts shared root causes, Project Zero ran dedicated experiments on bugs since 2023. Those experiments estimated true overlap at roughly 60%.
Either number leaves a large set of bugs that only one method finds. That is the argument for running both. Project Zero recommends fuzzing plus agentic analysis, not a swap.
Where LLMs Still Lose
Fuzzers excel at deep state and long input sequences. A JavaScript engine bug often needs thousands of generated operations to surface. LLMs are better at reading a diff and reasoning about intent.
Speed and cost differ too. A fuzzer farm burns CPU around the clock. BigSleep spends model tokens on one pass per day. The tradeoff favors LLMs for new code and fuzzers for hardened parsers.
Security teams should treat this as a coverage problem. Two tools with 60% overlap still cover more than either alone. Our October 25 security roundup covered why single-layer defense keeps failing.
Sources
Can You Run an AI Assistant With No Internet?
The Offline Emergency Kit
Survival LLM is a free, open-source app that runs an AI assistant and a reference library on your own computer. It answers questions about water purification, burns, and food safety during a blackout. Nothing you type leaves the machine.
The library includes medical references, FEMA and CDC guidance, survival manuals, repair guides, and about 930,000 Wikipedia articles. Every answer lists its sources, and you can open the exact passage with the cited text highlighted.
The full library is a 9 GB download and 19 GB installed. A smaller build without Wikipedia and iFixit is 1.3 GB, or 3 GB installed. Setup takes 20 minutes to about 2.5 hours.
Model Choice by Memory
The app picks a model based on your RAM. A Mac with 128 GB or more gets NVIDIA Nemotron 3.5 Super. A Mac with 32 to 96 GB gets Qwen3.8 27B. Both scored 21 of 22 on the project's blackout and first-aid test.
A Windows PC with 32 GB or more gets Nemotron 3.5 Lightning, which scored 19 of 22. Machines with 16 to 31 GB get IBM Granite 4.2 8B, which scored 14 of 22.
The scoring is honest about failure. The team graded answers blind against official guidance. Smaller models make more mistakes, so the app always shows the source.
Why It Matters Beyond Preppers
Local inference has quietly become practical. A 27B model that handles first-aid questions on a laptop was a research demo two years ago. Now it ships as a zip file with a source viewer.
The design choice is the interesting part. The app does not pretend the model is authoritative. It treats the model as a router to documents you can check yourself. That is a better pattern than a confident chatbot with no citations.
Sources
What Does Terence Tao Want From "Maths 2.0"?
Proof Scarcity Is Ending
Fields medallist Terence Tao published a Nature World View on October 22 calling for a "Maths 2.0" era. His argument is simple: AI has broken the bottleneck that defined mathematics for centuries. Producing proofs is no longer the scarce step.
Tao says AI-aided proofs have been released for hundreds of long-standing problems in recent months. Mathematicians may need months to understand the arguments and their implications. The field cannot absorb results at that rate using its current norms.
Three Problems With the Flood
First, many AI-generated results are not described clearly enough for publication. A proof that cannot be read is not a contribution, even if a checker accepts it.
Second, researchers move to the next problem once they have a minimally presentable proof. That leaves a trail of results nobody has verified or explained. The incentive rewards volume over understanding.
Third, attribution is unclear. Tao notes it remains unknown how much models draw on prior work by mathematicians without credit. The tradition of openly sharing problems is also under strain.
What Changes in Practice
Tao's proposal is cultural, not technical. Mathematics should prize human understanding over publishing more solutions. Journals, hiring committees, and funding agencies would all need to reward explanation and verification.
The Lean formalization wave makes this urgent. Machine-checked proofs are arriving faster than humans can state what they mean. Our October 27 analysis of OpenAI's ETH claim showed the same gap between a checked proof term and a understood theorem.
The parallel to software is direct. Generating code became cheap. Reviewing, specifying, and maintaining it did not. Every field that automates production eventually discovers that verification is the real job.
Sources
Frequently Asked Questions
Is Genie 4 available to the public?
No. It is a research preview for trusted testers. DeepMind plans access for Google AI Ultra subscribers in Project Genie and early partners on the Gemini API and AI Studio.
Why is Meta selling $95 billion in stock?
Meta says the money funds AI infrastructure and superintelligence work. It is also covering a capital spending plan of $135 billion to $145 billion this year alongside negative free cash flow.
Could an AI legally run a company in Delaware?
Not yet. Delaware has a draft bill for an Artificial Intelligence Company and a proposed 30-month sandbox. The bill has not passed, and a House committee is now requesting documents.
Does BigSleep replace traditional fuzzing?
No. BigSleep found 45 V8 vulnerabilities in six months versus 37 for Fuzzilli, with roughly 60% estimated overlap. Project Zero recommends running both approaches.
Can a small model handle emergency advice?
Partly. Survival LLM's 8B model scored 14 of 22 on the project's first-aid test, while 27B models scored 21 of 22. The app mitigates this by showing the source passage for every answer.
What Connects Today's Stories?
Capability, Capital, and Accountability
Genie 4 shows models generating whole sensory worlds. Meta shows the capital required to train them. Delaware shows the law trying to catch up.
The gap between capability and accountability is now the main story in AI. Every launch raises the same question. Who is responsible when the system acts?
For builders, the useful signals are concrete. Native audio in world models opens new interface work. Offline 27B models make local-first apps viable. LLM bug hunting works best as a second layer.
What to Watch Tomorrow
Watch whether the Delaware bill gets a hearing. Watch whether Meta's offering prices at a discount. Both will show how much risk the market and the state will absorb.