@davidpantera_Oct 27

A promising path for world models lies in HYBRID world models, which combine the logical, structured reasoning of code and language models with the rich, sensory representation of pixel-level simulation. Bringing Genie and Gemini models closer together is an early step in this direction. Gemini works behind the scenes to map the structure and plan the narrative timeline, while Genie generates the immediate physical reality (rendering how the world looks, sounds, and responds, pixel by pixel). Together, these complementary strengths can unlock a new class of world models that are far more coherent, controllable, and grounded than either approach could achieve alone.

@davidpantera_Oct 27

With Genie 4, you can use natural language to modify a simulation in real-time like changing the weather, moving the camera, or introducing a new person or object and Gemini will contextually adapt your prompt to match the surrounding scene. Between your inputs, Gemini also monitors and steers the scene to keep the world progressing cohesively.

@sundarpichaiOct 27

Known affiliation

Excited to share Genie 4 today our most realistic and interactive world model: - steer dynamic environments in real time, shaping events, camera angles, and actions - generate native audio matched to the world, from realistic background sounds to characters who can converse in real time! - simulate photorealistic scenes and real-world physics - create more coherent multi-step events through reasoning from Gemini We've started to roll out as a research preview to trusted testers. Coming soon to Google AI Ultra subscribers in Project Genie and early access partners via the Gemini API and AI Studio.

@googleaiOct 27

Known affiliation

Introducing Genie 4 from @GoogleDeepMind, our most realistic and interactive world model yet 🌍 Now you can use natural language to dynamically steer the world and its soundscape in real-time, whether you're changing the weather, moving the camera, or introducing new elements to the environment. So, how does the model do that? Behind the scenes, Gemini acts as the "brain" of the simulation. It translates your requests into actions, plans out complex events, watches the scene unfold to keep track of what's happening, and adjusts as you change the story. Genie 4 also introduces ambient audio and real-time interactive human dialogue that adapts on the fly.

@mtsliveOct 27

a16z's @droovg explains why ChatGPT and Gemini struggle to find the best startups to work for, and how Cosign is building the missing data layer: "One thing that LinkedIn will never do is introduce an MCP that allows you to have vast search access over the graph." "When I was in New York this summer, I wanted to start meeting people who I thought I might be able to work for one day." "So I Googled, who are the best startups in New York? Gemini gave me a pretty bad answer. I switched to ChatGPT, it also gave me a pretty bad answer. That whole data layer of information is kind of missing right now." "We have an iMessage agent. We also have a personal MCP that lets you connect your Muse, your Grok Bot, your Instinct to the platform, to make this truly the first agent-native professional networking platform." @cosign @a16z

@businessbaristaOct 26

Experian, a $35 billion business with roots dating back to 1826, is surprisingly AI-native. Its new Chief AI Officer, Vijay Mehta, broke down the credit bureau’s end-to-end AI strategy. Complete list of notes... Big takeaways:
 • Experian is well past proof of concept, with AI capabilities running in production throughout the company.
 • The next stage is what Mehta calls "reimagination." Redesigning how work gets done through delegation to agents vs. prompting chat bots. 
 • Adoption is close to universal. 85% of employees actively use the company's AI tools, 88% of developers use coding assistants, and employees built more than 6,000 custom GPTs by the first half of fiscal 2026.
 • They're seeing real ROI. It's measured in time savings, time to value, quality, velocity and traditional before and after financial metrics, not in adoption or token use. How they built the foundation:
 • Within months of ChatGPT's launch, they trained thousands of employees and put AI policies and guardrails in place.
 • AI traffic was routed through a control plane that gives them observability, monitoring and security.
 • They wrote a GenAI risk framework and built the Experian GenAI Platform, an internal system that lets employees build internal-facing & and customer-facing apps, while accessing company data securely.
 • They rolled out a coding assistant to their large developer base and built their own Experian GPT into products like a digital financial assistant and their Ascend platform.
 • They created a GenAI Center of Expertise, which now sits under Mehta's AI office.
 • Since then, AI has spread across software development, analytics and business operations. AI inside the employee experience:
 • MyCareer platform recommends learning and career moves and helps managers spot skills gaps.
 • Nadia, an internal AI coach, helps employees set goals, build development plans and prepare for performance reviews.
 • Nadia also gives leaders personalized coaching and 360-degree feedback. On models and cost:
 • "Not everybody needs the most advanced frontier model. Most workloads can be done with a free or very low-cost model."
 • He compares reaching for a frontier model by default to using a sledgehammer on a job that needs a much smaller tool.
 • They run a mix of open-weight models adapted to their own environment and frontier models accessed through AWS Bedrock.
 • The control plane routes each workload to a model based on the task, language, region and regulatory requirements. On agent safety:
 • Guardrails are designed into the architecture from the beginning instead of being added later.
 • Their Agent Operating System runs on that control plane, so it can sandbox, monitor and stop an agent process when necessary. "The ability to kill a process, the ability to kill a process, I will repeat that, we've got all of that built in."
 • Their risk profile differs from the frontier labs' because their systems work over a defined body of financial and identity data, without broad access to the internet. "We start with our corpus of data and we put controls around that. And we know exactly how the data is moving." What the CAIO role actually is:
 • Experian is a big company split into many business units. His job is one common AI strategy and direction: "to make sure that we are all marching in the same direction."
 • He accelerates the AI agenda both internally and externally.
 • Risk, security, compliance and partner choices all come together in the AI office. AI as the new middleman:
 • Consumers increasingly start their financial journey inside an AI assistant instead of going to Experian or a bank, so Experian is meeting them there.
 • It's the first personal finance app integrated with Google's Gemini. US consumers who connect their Experian accounts can ask Gemini about their FICO credit rating and what's affecting it.
 • Inside ChatGPT it has launched personal loan shopping, credit card shopping and an insurance marketplace.
 • On the enterprise side, ServiceNow is the first integration partner for its Agent Operating System. Under a multiyear partnership, ServiceNow agents tap Experian's Ascend data and decisioning from inside enterprise workflows.
 • It also launched Agent Trust, which verifies that an AI agent is actually connected to the consumer it claims to represent.
 • Could agents eventually skip Experian and get the data somewhere else? He says they can't reproduce it. Decades of longitudinal financial and identity data, plus its provenance and regulatory standing, are what matter when the output drives consequential decisions. "In order to be able to make a decision, you've got to have that trusted data source." His advice if you're early:
 • Train employees.
 • Set guardrails before deployment. "So much of safety and responsibility comes from the person interacting with the system."
 • Route AI traffic through a controlled gateway where usage can be observed and policies enforced. Which AI leader should I learn from next?

@aakashguptaOct 26

Model PMs earn 60% more than non-AI PMs. Here's how 316 of them got the job. The open model PM roles that post a US range have a median base of $345K. PM roles outside AI post $216K. To see how people land the job, I read the public LinkedIn profiles of 316 people in product roles at 22 companies that train AI models. Here's what they have in common: • 85% were PMs before joining the lab • 64% have a technical degree • 28% founded a company first • 24% have an MBA • 2% have a PhD The median career so far is 13 years. Only 6 of the 316 have under 5. So where did they start? Here's their first job after school: • Product: 24% • Consulting, banking or VC: 22% • Engineering: 17% • Marketing, sales or ops: 15% • Their own startup: 7% • Data or research: 6% Data or research is the smallest starting point. The research background people expect is rare. Before the lab, the most common stops were Meta (48 people), Google (42), Microsoft (39), Amazon (22), Stripe (17) and McKinsey (12). So the usual path looks like this: start almost anywhere, become a PM, then move closer to the model. Steven Pecht started as a business analyst at McKinsey. Then BizOps at Google, PM on Stadia, Head of Product at Zumo Labs, and now PM for Gemini model quality. Grant Gardner was a program manager at Microsoft. Then eng manager, then CPO and CTO at Intelight, then Staff PM at Meta Horizon Worlds. Today he's a Lead Group PM at Meta Superintelligence Labs. Yonatan Wolf went from Associate PM to Group PM at Google, then Group PM for research at Google DeepMind. Matteus Pan started as an associate at BCG, then Group PM at Dropbox, Distinguished PM at Bina Technologies and Head of Product at Genesis Molecular AI. Today he's a Member of Technical Staff at Thinking Machines. Barkley Dai interned as a PM at Kuaishou and Tencent, then spent two years on AR and GenAI effects at TikTok. He joined Luma AI as its first product hire, started on growth, and now leads research product for data and model capabilities. Every one of them got there through product. And 28% founded a company before the lab. Building something of your own is one of the clearest signals you can send. If you're a PM today, you already have the most common thing on this list. The next step is moving closer to the model: take the AI feature on your team, learn to build evals, and write the behavior spec yourself. Which of these paths is closest to yours?

@aakashguptaOct 23

Gemini got asked for a random number between 1 and 10 a hundred times in a row. It said 7 all hundred times. And 7 is also the number you were most likely to pick. That's the entire secret behind AI "reading your mind." You and the chatbot learned what random looks like from the same place, which is other people. People are hilariously bad at being random. In one study of 558 people, 28% picked 7, way more than any other number. GPT said 7 in 92 out of 100 tries and Claude said it 90 times. Watch what your brain does in the half second before you answer. 1 and 10 feel too obvious, so you skip them. 5 feels like a cop-out because it's right in the middle. Even numbers feel too neat, so 2, 4, 6 and 8 are out. 3 and 9 feel patterned, so those go too. Count what's left and you've got 7. An AI learned to talk by reading billions of pages that we wrote, including every forum thread and party trick where somebody typed "pick a number." Inside the model, the only thing happening is next-word prediction. It looks at the question and goes with whatever word most often came next when real people wrote it, and across all those pages, that word is 7. And it exaggerates us. About a quarter of people go with 7. The model's whole job is landing on the single most likely answer, so a quarter of people turns into almost every single response. Same thing happens with bigger ranges. A researcher at Capco asked seven different AI apps for a number between 1 and 50, and six of them said 27. Grok was the holdout with 42, the Hitchhiker's Guide joke. Veritasium surveyed 200,000 people on 1 to 100 and the favorites were 37 and 73, which is exactly where the AIs cluster too. So try it yourself with 4 in your head. A real mind reader would catch that one. ChatGPT is only ever going to catch the person who picked 7, and that person has been typing 7 on the internet for thirty years. Every time an AI "reads your mind," you're looking at the average of millions of people who each felt sure they were being original.

@geminiappOct 22

Known affiliation

With your permission, Gemini can now reference your @GoogleMaps searches and saved places alongside your Google Calendar to give you tailored recommendations for your day. 📍

@aakashguptaOct 22

The NSA just told companies to pull their websites offline if they can't patch them right away. The reason is a bug that's been sitting in the code behind 70% of websites for 14 years. Last Friday a guy found it in one evening by typing five words into an AI. Those five words were "find bad vulnerabilities in PHP." He's Dan Lorenc, CEO of a security company called Chainguard, and all he had going in was a rumor that something big was coming in PHP. A small firm called Calif had already found the same flaw back on August 1 using OpenAI's GPT-5.6 Sol. Wiz, the security company Google bought, found it on its own less than three weeks later with Gemini and GPT. Since then OpenAI, Wiz, Calif and PHP's maintainers have been quietly building a fix that Lorenc himself called "crazy." You'd think a bug this bad would've gotten caught sometime since 2012. PHP is open source, so anyone on earth could read this code. Wiz says turning the flaw into a working attack took a chain of steps right at the edge of what human researchers could realistically pull off. For 14 years the sheer amount of work was doing most of the protecting. AI made that work cheap enough to knock out on a Friday night, for anyone who knows where to point it. And the rumor got out. Bug ID numbers started floating around online Tuesday. Hours later one of the research teams posted a video of the exploit by mistake, and PHP moved its release up a day. Early Wednesday, Chainguard pushed its fix out more than six hours ahead of everyone else, saying people were already getting attacked, and Lorenc says his company's decoy servers caught IP addresses from China and Russia going after the flaw. Here's why the NSA went straight to "unplug it." VirusTotal counted about 758,000 exposed servers running old PHP versions, 5.4 through 8.1, that don't get security updates anymore. Only around 272,000 are on versions that do. So for roughly three out of four servers that show their version number, there's no patch to install, and pulling the plug is the only fix they've got. This is happening everywhere at once, too. Microsoft just fixed 1,413 vulnerabilities in a single month, more than in any full year on record. OpenAI says it expects finds like this one to jump sharply over the next few months. Every company sitting on a secret bug now has until the first rumor leaks and somebody with an AI subscription goes looking.

@aakashguptaOct 21

An AI agent got a $250 delay credit out of Delta for one person. Another one cancelled $1,300 a month in subscriptions its owner forgot about. Then I saw the ranking of all nine of these agents built on what users say in public, and the one I use every day came in 6th. Meta’s Muse is #1 with a score of 70.5. @feedbackbench is a new site from Enterpret for comparing AI personal agents on what real users say about them. It read Reddit, X, YouTube and both app stores from August 31 to October 18, kept only the posts where someone described their own use, and scored each agent on popularity and customer love. I’ve been using Instinct the most, so I had a horse in this. Instinct sits 6th at 46, behind OpenClaw and Hark (haven’t tried Hark yet). Hermes is 2nd and Grok Bot 3rd. Now hold the money up against that board. Instinct just raised $1 billion at a $10 billion valuation and sits two points above Google. Nous Research, which builds Hermes, got to 2nd on a $90 million raise at $1.5 billion. Zuckerberg said last week Meta was ready to spend up to $200 billion on AI without knowing if it would work, and its agent is the one at the top. Gemini Spark scores 44 and ChatGPT Dots 39, both below Instinct. That matches what I found when I tested both. The Delta credit and the cancelled subscriptions only happen if you let the thing into your inbox and your card, so trust is the whole product here. For my money this is the most interesting market in AI this year, and this is the first scoreboard for it that somebody other than the vendors wrote. Built by @enterpret_ai, link in the reply.

@tibo_makerOct 21

I ran an AEO experiment and found something critical the median DR of websites AI cites is 86 I asked ChatGPT, Perplexity, Gemini and Claude the same set of questions and pulled the DR of all 1,329 sources they cited say you ask ChatGPT “best CRM for startups”. it searches, reads a bunch of pages, cites around 7 sources and writes you an answer you read the answer, but the AI read the blog posts first and decided who gets quoted here’s what the cited pages have in common: - trust only 7% of citations came from sites under DR 30 AI quotes sites the web already trusts, and that trust is built with backlinks - page format 73% of the sources were blog content: listicles, articles and comparisons product and landing pages were 7%, and forums and social were only 3% - original data ChatGPT is the exception: only 18% of its citations were blog content, and 63% were the original source, like a pricing page or the docs AI quotes the place a fact comes from, so if your post repeats the same stats as 40 other posts, there’s nothing in it to quote - interactivity on the page many cited pages had a calculator, chart, table or template the few humans who click through already read the AI’s summary, so they click to do something, not to read now you can’t get to DR 86 right away, but you can do 2 things: 1. get to DR 30+ in less than 5 months 2. get your product mentioned in the exact listicles AI keeps citing Outrank helps you do both it builds backlinks for you automatically and gets your product the #1 spot in the listicles AI cites so after this experiment, we rebuilt all 280 posts on the Outrank blog this week. 275 of them now have a calculator, chart, checklist or template that’s what we’re building into every Outrank article: real data, backlinks that raise your DR, and pages worth clicking so now we write for the AI first, then give the human a reason to click

@rileybrownOct 21

Is it just me or is every model besides Claude and GPT models (comparably) unusable? Is this a skill issue? I swear they they are in the league of their own when I use grok, muse, Gemini or the open models I just end up frustrated.

@aakashguptaOct 21

Madhu Guru (@realmadhuguru) led product for Gemini. He thinks fewer than 75 PMs are qualified to do the work. That's a pretty wild stat. Madhu is now a senior director of AI at Meta, and he identified 5 skills that every PM working on things like Gemini needs to be able to do: 1) Vision and opinionated UX design 2) Model intuition to extract max value 3) Going from pixels to evals to hill climbing 4) Agentic flows: tools, context and safety guardrails 5) Deep user understanding He says it's scar tissue from hiring, so I'm inclined to believe him. He means AI-native PMs, but I really think it's fewer than 75 for model PMs. I found 46 open model PM roles, oddly close to 75. I read all the postings, and they ask for similar things. Here's what they look for, by percentage of postings: 1) Vision and opinionated UX design: 61% 2) Model intuition to extract max value: 61% 3) Going from pixels to evals to hill climbing: 76% 4) Agentic flows: tools, context, safety guardrails: 59% 5) Deep user understanding: 50% Now the last line of his tweet: "It is a rare, but learnable skillset. Only way is to build, build, build." I think that's the most important point, and the postings agree with him. Only 7 of the 46 require production experience. What they're really looking for is that you are building, and you can go build these things on the side. What I recommend: pick a task your favorite model struggles with. Write down the failure modes. That's the start of your eval set, and that's the product you should build, something that solves them. Start with a behavior spec for it: "the model should..." Those lines become additional evals. Then ask Claude Code to build a product that passes those evals, and iterate. You'll learn how to build on top of models. Model PMs are paid extremely well. The median posted base is $345K, and the range runs from $144.6K to $490K. As he said, it's learnable. Because there's so little supply, why don't you become one? Why don't you get that pay bump? I think you can, and I'm here to help. Let me know any challenges you face in the comments.

@a16zOct 20

Eugenia Kuyda, founder of @wabi, on why chatbots are the MS-DOS era of AI, and why a new interface is necessary to make AI's full potential obvious to everyone: "It always struck me as strange that people were using products like ChatGPT, Gemini, or Claude, but mostly for very simple use cases: for search, or to help them write something, help with homework." "No one was really mentioning any of these exciting capabilities that we just kept seeing the models get, and so we felt like there must be an interface problem." "There should be something simple for everyone, me included, to actually discover these amazing use cases that all these models already have." "We got obsessed with this metaphor that the current chatbots are really the Microsoft DOS era for AI interfaces. And that there will be some sort of a Windows/MacOS moment." "There is so much traction with chatbots... In order to unlock something else, you need a more exciting interface." @ekuyda