Welcome back! A new report just put numbers on something most of us already suspected: Google's AI Overviews are confidently wrong at a staggering scale. Meta finally shipped the first model from its Superintelligence Labs, Elon Musk revealed that xAI is training seven models simultaneously on Colossus 2 (including systems with up to 10 trillion parameters) and Perplexity is so confident in its agent tools that it is betting $1M you can build a billion-dollar company with them.
In today’s Generative AI Newsletter:
Google AI Overviews: How many wrong answers per hour is acceptable when you control 90% of global search?
Meta Muse Spark: What did Alexandr Wang actually build with nine months and a blank cheque?
xAI Colossus 2: What does training seven models at once, including a 10-trillion parameter system, tell you about where Musk is heading?
Perplexity: Revenue jumped 50% in a single month. Now the company is giving $1M to anyone who can build a unicorn with its tools.
Latest Developments
Google's AI Is Confidently Wrong Millions of Times Per Hour
A New York Times investigation found that Google's AI Overviews are accurate roughly 90% of the time. That sounds reassuring until you consider the volume. Google processes around 5 trillion searches per year. A 10% error rate at that scale means tens of millions of overviews per hour are presenting wrong information with the same confident formatting as correct answers.
The details:
Why It Fails: The study found three recurring causes. Overviews frequently pull from user-generated content on Facebook and Reddit (the second and fourth most cited sources). They link to websites that do not actually support the claims being made. And they generate false summaries of otherwise factual source material.
How Easy It Is to Game: Journalist Thomas Germain demonstrated the vulnerability by publishing a blog post titled "The Best Tech Journalists at Eating Hot Dogs," ranking himself first. Google's AI served it up as fact.
The Trust Problem: Every AI Overview carries the same visual authority regardless of whether it is correct. There is no confidence indicator, no margin-of-error disclosure and no easy way for the average user to tell a verified answer from a hallucinated one.
The Scale: At 5 trillion annual searches, even a 1% error rate would mean billions of wrong answers per year. At 10%, the numbers are difficult to process. And that 90% figure is the overall average. Accuracy on medical, legal and financial queries, where errors carry the most consequence, was not broken out separately.
Google's AI Overviews sit at the top of search results, above organic links, above ads and above every other source. They are the first and often the only thing people read. For the millions of users who treat that answer box as gospel, 90% accuracy at 5 trillion searches per year is a liability dressed up as a feature.
Special highlight from our network
Right now, a lot of people feel this pressure: AI is everywhere, but they’re still not quite sure how to use it. It’s all about the fear of falling behind.
This fear is completely justified. The gap is growing between people who can apply AI in their work and those who can’t. And that gap shows up in speed, opportunities and confidence.
This 2-day AI Mastermind from Outskill is designed to help you start using AI in your day-to-day work (because understanding is never really enough).
What will you get out of it?
Clarity: what tools matter and where to start
Practical skills: writing, automation, basic AI flows
Efficiency: ways to save time in your day-to-day tasks
Career edge: understanding how AI can make you more valuable
You’ll also receive templates, toolkits and guides to help you apply what you learn faster.
Dates: April 11–12 | 10 AM – 7 PM (EST) | Online | Free
Register now
Meta's Superintelligence Lab Just Shipped Its First Model
Meta rolled out Muse Spark, the debut release from Alexandr Wang and Meta Superintelligence Labs. Wang took over the division nine months ago after Zuckerberg acquired Scale AI for $14.3B and says the team rebuilt Meta's entire AI stack from scratch to get here.
The details:
What It Does: Muse Spark handles voice, text and image inputs. A "contemplating mode" pits multiple internal agents against each other on hard problems before surfacing a final answer.
Where It Stands: Benchmarks are competitive with Opus 4.6 and GPT-5.4 on reasoning, though it falls behind on coding and tests like ARC-AGI 2.
Health Focus: The model is particularly strong in health reasoning. Meta is prioritising medical applications as part of its broader "personal superintelligence" mission.
Proprietary Shift: Unlike the Llama family, Muse Spark is proprietary. Meta says it hopes to open-source future Muse models but has given no timeline.
Access: Available now in the Meta AI app and website, with API access for selected partners. Rolling out across Meta's other platforms in the coming weeks.
Muse Spark is competitive without being frontier-leading. That matters less than what comes next. Meta has 3 billion daily users, more consumer data than any other company on earth and functionally unlimited compute budget. Wang spent nine months rebuilding the stack. Everything that ships from this lab next will build on what landed today.
xAI Has 7 New Models Training at Once
xAI is currently training seven new models simultaneously on its Colossus 2 supercomputer, including systems with 6 trillion and 10 trillion parameters.
The details:
The Hardware: Colossus 2 is xAI's expanded supercomputer cluster. The original Colossus was already one of the largest AI training installations in the world. Version 2 has the capacity to run multiple frontier-scale training runs in parallel.
The Scale: A 10-trillion parameter model would be among the largest ever trained. For reference, GPT-4 was reportedly around 1.8 trillion parameters. Anthropic's Mythos, revealed last week, was described internally as a step change. xAI appears to be aiming for a similar leap.
The Strategy: Training seven models in parallel suggests xAI is running multiple architectural experiments at once rather than betting everything on a single approach. That is expensive but it compresses the timeline for finding what works.
The Context: This comes alongside Elon Musk's amended lawsuit against OpenAI (redirecting damages to the nonprofit, pushing to remove Altman from the board) and Intel's announcement that it is joining the Terafab project alongside SpaceX and Tesla.
Musk has spent the past year building the infrastructure. Colossus, the Memphis data centre, the Terafab chip partnership with Intel. The model that justifies it all is still missing. Seven simultaneous training runs suggest that is about to change, though whether any of them will match the frontier remains an open question.
Perplexity Will Pay You $1M to Build a Unicorn
Perplexity's annualised recurring revenue hit $450M in March after jumping 50% in a single month, according to the Financial Times. The company now has over 100 million monthly active users. And it is so confident in its own momentum that it just launched the Billion Dollar Build: an 8-week competition offering $1M in seed funding to anyone who can build a company with realistic unicorn potential using Perplexity Computer.
The details:
The Revenue: ARR grew from $305M to $450M in one month. The jump was driven by Computer (its agent-style tool that orchestrates up to 19 models from OpenAI, Anthropic and Google) and a shift to usage-based pricing.
The Competition: Register by April 14. Build a real company using Perplexity Computer over 8 weeks. Perplexity selects 10 finalists to present live. Up to 3 winners split $1M in seed investment and $1M in Computer credits.
The Pivot: Perplexity started as an AI-powered search engine. It is now positioning itself as a platform for building businesses. The competition formalises that shift and doubles as a customer acquisition strategy for Computer.
The Wider Picture: Perplexity's growth still trails the biggest AI startups. Cursor hit $2B in ARR. Anthropic reported $19B at the end of February. But $450M from a company that was generating $16M two years ago is a growth curve that demands attention.
The competition is smart. It turns users into case studies, generates PR and stress-tests the product at the same time. If even one winner builds something real, Perplexity gets a portfolio company and a proof point. If none do, the investment was still cheaper than most advertising campaigns.
Tool of the Day: Wispr Flow
Did you love the dictation feature in ChatGPT? But hate the same function in your iPhone? You’re going to love Wispr Flow. This tool brings elite level voice recognition to your Mac, PC, iPhone and Android. Understanding levels and text replacements are intelligent, ensuring very high accuracy and speed.
Try this yourself:
Open any app where you type, tap the floating bubble and start talking.
Use it for long emails, detailed prompts, Slack messages or code comments.
It works anywhere you can type. It’s typing these words right now.
Light Bytes
Anthropic launched Claude Managed Agents in public beta. Define tasks, tools and guardrails. Anthropic handles the infrastructure. Agents can run solo for hours at $0.08/hr on top of usage fees. Notion, Rakuten, Asana and Sentry are early adopters.
HeyGen released Avatar V, a model that builds a full video avatar from a 15-second phone recording and claims to have eliminated identity drift (the tendency for AI-generated faces to stop looking like you over time).
Meta employees burned through 60 trillion tokens in 30 days. An internal leaderboard called Claudeonmics tracked the top 250 users, with the leader racking up 281 billion tokens. Meta shut the tracker down after the data leaked externally.
Elon Musk amended his lawsuit against OpenAI, redirecting all damages to the nonprofit arm and pushing to remove Sam Altman from its board. OpenAI called it a harassment campaign.
Canva acquired Simtheory and Ortto, adding agentic AI workspace tools and marketing automation as it pushes toward end-to-end campaign management. Blank canvas.
