Google just made a game engine you can talk to
Genie 4 generates explorable 3D worlds in real time, with physics, sound and characters you can talk to. You can change the weather, add objects or people, and reshape the world while playing. Some of the demos look surprisingly close to actual games.
Imagine an MMO where players can build towns, invent quests or create entire worlds just by describing them. No game editor, no coding. I’d probably spend more time making things than actually playing.
There’s still the small problem of objects disappearing, physics going wonky and worlds losing consistency over longer sessions. But Google is already testing Genie with game developers, filmmakers and robotics teams. It’s in limited research preview, with AI Ultra access coming.
Google is back at #1 with Gemini 4 Carbon
A day after Genie 4, Google released another frontier model. Carbon scores 59 on Artificial Analysis’s Intelligence Index, narrowly beating Claude Opus 5.5 (58) and GPT-6 Astra (57). Google says its engineers have had it working on difficult problems for days, recovering from failures and fixing its own mistakes. Would love to see that tested independently.
But Sundar Pichai’s comments on Google’s earnings call might be the bigger story. He says AI is already helping Google train and release models faster, and expects further acceleration over the next 12 months because of recursive self-improvement (RSI). AI helping build better AI, in other words.
That doesn’t mean Google has cracked self-improvement, but it’s quite something to hear its CEO talking about it so openly. Argon became widely available just 15 days ago, and we’re already discussing its successor. If AI is genuinely speeding up model development, I wonder how long a flagship model will even remain a flagship.
Carbon is limited to selected cybersecurity defenders for now.
Anthropic is coming for the legal department
Claude for Legal now has ready-made tools for reviewing contracts, monitoring regulations and handling legal requests. Teams can customize and share them, with links back to the original sources.
Anthropic already has tools for finance, accounting, design and spreadsheets. Now legal. At some point you start wondering which department isn’t getting its own Claude.
Harvey and Legora have raised hundreds of millions building specialized tools for lawyers, and now the companies supplying the underlying intelligence are starting to build the workflows too. A lot of expensive legal software could eventually become a feature inside a Claude subscription.
There’s a funny twist, though. Harvey actually helped build one of the new Claude tools, and Legora is an integration partner. So they’re simultaneously customers, partners and potential competitors. Not an easy relationship to navigate.
I don’t think legal AI startups are doomed, though. The ones deeply embedded in law firms, with trusted data and workflows, have a real advantage. But charging thousands for something a frontier model can eventually do for a fraction of the price is becoming a risky business.
Anthropic ships Claude for Legal with ready-made contract review and compliance tools
Claude for Legal just got a lot more useful. Instead of typing questions into a chat box, you now get a gallery of pre-built interactive tools called artifacts, built specifically for legal work.
Think of artifacts as ready-to-run apps living inside Claude. You pick one, load your own documents, and Claude does a first pass. You review and decide. That’s the whole loop.
Six workflows ship ready to use:
- Cite Check: flags misquotes, wrong page numbers, and citations that don’t back up the claim
- Tabular Contract Review: answers questions across a stack of contracts and shows the exact source quote
- Regulatory Scan: monitors for new legal developments and routes them to the right owner
- Playbook Deviation, Jurisdiction Matrix, and Matter Board round out the set
You also get 60+ connectors to tools like Bloomberg Law, Docusign, and Filevine. Firms like Goodwin and Foley and Lardner contributed their own artifacts too.
The big unlock: finished workflows you can share across your whole org, no building from scratch required.
Liquid AI ships six open-weight 2B models that run on-device with 35% fewer parameters
Liquid AI just dropped LFM3, a family of six small models that run on your laptop, phone, or Raspberry Pi. The big idea: tiny models that punch way above their weight.
The base model, LFM3-1.7B, beats a 2.6B model from their last generation using 35% fewer parameters. Its weights fit in about 1GB, and it handles a 256K-token context window, meaning it can hold a massive amount of conversation history without choking on memory.
Here is what the six models cover:
- Text with long context and agent tool use
- Vision and video understanding, competitive with models 6x larger
- Audio with 18-language voice and tool calls
- Omni combining vision, audio, and language together
- Two decision models that output structured, typed answers in one pass
You can fine-tune any of them on a single GPU, then deploy locally with no API fees. Weights are open on Hugging Face, and it runs on Ollama, vLLM, and llama.cpp today.
Artificial Analysis drops v5 benchmark index with Claude Opus 5.5 leading at 58
AI benchmarks have a cheating problem. Labs train on public test data, scores inflate, and rankings stop meaning anything. Artificial Analysis just fixed that with Intelligence Index v5.
The big move: 60% of the index now uses private test sets, meaning models can’t be trained on the answers ahead of time. It’s a direct response to benchmark contamination.
Two new benchmarks you should know:
- FrontierCode 2.0: 132 real open-source repo tasks, graded by whether a maintainer would actually merge the code. No partial credit.
- Terminal-Bench-Science: 70 tasks where an AI agent runs real scientific research workflows inside a terminal. Every single test must pass to count.
Current standings: Claude Opus 5.5 leads at 58, GPT-6 Astra sits at 57. The open-source gap is the real story though. It jumped from 3 points to 12 points behind the leader. Downloadable models now score 26 points below the top.
Cost matters too. GPT-6.1 Sol scores within 4 points of Opus 5.5 at roughly one-eighth the cost per task. Effort settings now swing scores by up to 23 points.
Carbon Couldn’t Save Google
On Wednesday Google released Gemini 4 Carbon, its “strongest model,” and it took first place on one independent index. That evening Sundar Pichai put it in his headline earnings quote: “Gemini 4 Argon and Carbon are now leading at the frontier.”
Investors were not following. Alphabet shares fell more than 10% after hours, erasing over $280 billion in market value. The company missed forecasts on revenue and earnings for the first time in nearly four years, and Search grew 12%, its slowest pace this year.
The problem is the bill. Capital spending reached $51.4 billion, more than double a year earlier, and free cash flow was negative $9.1 billion. Finance chief Anat Ashkenazi said capex will “increase significantly in 2027.”
Carbon cannot pay any of that down soon. This week only “a subset of cyber defenders” get access, and developers, businesses and consumers wait while Google iterates “on guardrails.” Carbon and the selloff landed on the same day. Only the selloff was open to the public.
Musk Buys the Neighbors
Elon Musk’s AI company has built three data centers around Memphis and Southaven, Mississippi, in two years by treating every obstacle as optional. Power came first: 69 gas turbines, run without the permits federal law requires.
The neighbors came next. After residents sued over the noise, agents for the SpaceXAI unit that runs the plant began going door to door in Colonial Hills: market value for your home if you sign a release of all claims, class action included.
The target is 418 homes, a realtor told one resident, nearly the whole neighborhood. At local prices that is roughly $75 million, small change for a data center buildout.
City, state and Washington have each taken its side. Southaven’s mayor calls the buyout “a goodwill gesture from a good corporate citizen.” Mississippi’s governor says curbs on data centers mean “surrendering the future of technology to China.” When the NAACP sued over the turbines, the Justice Department stepped in on SpaceXAI’s side, citing national security.
The mayor says the buyout “is not intended to be a company city at this point.” A fourth data center, called Minihard, is already under way.
Human Review Lost to a Coin
Anthropic sells Claude to professionals on one promise: the human still makes the call. Claude for Legal, its product for lawyers, says every output is “a draft for attorney review” and that “a lawyer reviews, verifies, and decides.”
This week Anthropic tested what it calls the usual assumption, that experts “still make the decisions that matter.” It gave 50 professional analysts a statistics job in Claude Code and let a coin flip secretly decide which option Claude would argue for on four judgment calls. The analysts went with Claude 86% of the time, against about 50% by chance.
They did not notice. Afterward, 77% defended their choices with Claude’s arguments, sometimes word for word, and only 15% credited it. Analysts who followed Claude on every call were as confident as the rest. Anthropic says it strengthened the leans on purpose and tested only one task.
The coin flips alone predicted 82% of the analysts’ final recommendations before they had started. The expert still reviewed, verified and decided. The coin had decided first.
GitHub Is Finally Fixing Itself
GitHub is finally willing to fix itself. At its Universe conference on Wednesday it said it has added 4 million CPU cores this year, is redesigning Git to split storage from compute, and will finish moving to a multi-region Azure setup early next year. “Capacity alone won’t get us there,” wrote Jay Parikh, the Microsoft executive over GitHub.
AI agents forced the issue. Monthly commits went from 1.4 billion in April to 7.4 billion in September, and an outage on August 17 ran for nearly eight hours. “We let you down,” its CTO wrote.
Rivals did not wait. That same day, SpaceX-owned Cursor launched Origin, a competing code host. Roku moved its developer repositories to GitLab this month. This week Entire, a startup run by former GitHub CEO Thomas Dohmke, opened its own GitHub alternative to everyone, pitched at the outages.
What still holds customers is not uptime. “The lock-in is Actions,” one engineer said in August. The fix has to land before they rewrite them.