In partnership with
Welcome to OHMO AI. Moonshot dropped a 2.8 trillion parameter model and put the weights up for anyone to take. TikTok started scanning its own platform for fake versions of your face. And Bezos quietly backed a Cambridge startup trying to invent materials that don't exist yet.
In today's newsletter:
Moonshot's Kimi K3 lands as the largest open model ever released
TikTok tests a tool that hunts down AI clones of creators
Bezos and the UK government back CuspAI at a $2.6 billion valuation
OpenAI's CFO says the industry is measuring AI wrong
LATEST IN AI
MOONSHOT AI
Kimi K3 Puts A 2.8T Model In The Open
Image source: Kimi
OHMO AI: Moonshot released Kimi K3, a 2.8 trillion parameter model with native vision and a 1 million token context window, calling it the world's first open 3T class model. The company says full model weights will be released by July 27, and it's already live on Kimi.com, Kimi Code, and the Kimi API.
The details:
Moonshot is upfront that K3 still trails the top proprietary models, naming Claude Fable 5 and GPT 5.6 Sol as ahead of it.
The company claims roughly 2.5 times better scaling efficiency than Kimi K2, driven by a new attention architecture and a sparse expert setup that activates 16 of 896 experts.
In one internal test Moonshot says K3 autonomously designed a working chip in a single 48 hour run. That figure is the company's own, unverified by outside labs.
API pricing is $3.00 per million input tokens on a cache miss and $15.00 per million output tokens.
Why it matters: Every benchmark here comes from Moonshot, so treat the numbers with care. What isn't in dispute is the release itself. A model this size going open means anyone with enough hardware gets frontier-adjacent capability without asking permission from a US lab, and that's a different world than the one where three companies hold the keys.
TOGETHER WITH OYSTER
What's changing in global hiring?
Global hiring is changing fast. From AI's impact on HR to evolving compliance requirements and international expansion strategies, the rules are constantly shifting.
Oyster's events bring together HR leaders, founders, operators, and global employment experts to discuss what's working now—and what's coming next.
Whether you're hiring internationally today or planning for tomorrow, you'll walk away with practical insights you can actually use.
TIKTOK
TikTok Starts Hunting Your AI Double
Image source: OHMO AI
OHMO AI: TikTok is testing an opt-in Likeness Detection tool with a small group of US creators, scanning the platform for AI-generated video that uses their face without consent. The feature was first spotted by social media consultant Matt Navarra, and TikTok spokesperson Zachary Kizer confirmed the limited test.
The details:
Creators who opt in verify identity through Jumio with a live selfie and a government ID check.
TikTok says it doesn't keep copies of ID documents and uses facial data only for likeness matching.
Once enrolled, the tool flags likely matches and lets the creator report the video or the account behind it.
YouTube got here first, expanding its own likeness detection through 2025 and 2026.
Why it matters: The trade is right there in the setup. To stop someone from faking your face, you first hand over a scan of your real one. TikTok says it won't keep the ID, and it's opt-in rather than Meta's default-on Muse Image approach that got pulled after three days. Whether "opt in" stays optional once deepfakes get cheaper is the part worth watching.
CUSPAI
Bezos Backs A Search Engine For Matter
Image source: OHMO AI
OHMO AI: Cambridge startup CuspAI closed a $450 million round at a $2.6 billion valuation, with money from Jeff Bezos's family office Bezos Expeditions and Britain's Sovereign AI Venture Fund. Alongside it the company launched the AI Materials Foundry, a coalition of more than 45 partners including Nvidia, Meta, and Hyundai.
The details:
CuspAI's platform works like a search engine for materials. You describe the properties you need, it proposes chemical structures and synthesis routes.
One stated goal is cutting or removing rare metals like iridium and ruthenium from chipmaker supply chains.
The round was led by Kleiner Perkins and NEA. It's up from a $520 million valuation last September.
CuspAI is the fourth company backed by the UK's sovereign AI fund, which typically puts in £1 million to £10 million. The government didn't disclose its stake.
Why it matters: Most AI money chases software that makes more software. This is a bet that the bottleneck for batteries, chips, and carbon capture is physical, and that the thing standing in the way is a material nobody has found yet. Founders including Geoffrey Hinton and Yann LeCun sit on the advisory board, which tells you where some of the field thinks the next decade goes.
PRESENTED BY VIKTOR
Hampton took $440K in planned hires off the calendar
Hampton co-founder Joe Speiser had three roles budgeted: a data engineer, an ops manager, a PM. $440K. He installed Viktor on April 12. Forty-four days later, none are on the calendar, and 18 of his team work with Viktor daily. His VP: we are editors now, not creators.
OPENAI
OpenAI's CFO Says You're Measuring Wrong
Image source: OHMO AI
OHMO AI: OpenAI CFO Sarah Friar published a framework arguing companies are tracking the wrong AI numbers. Her pitch is a metric she calls "Useful Intelligence per Dollar," built on four questions: how much real work got done, what each successful task cost, how often the result was usable, and whether the economics improve at scale.
The details:
Friar argues cost per token is misleading, since a cheap model that needs three attempts can cost more than an expensive one that gets it right first time.
She points to GPT-5.6's three tiers, Sol, Terra, and Luna, as a way to match model cost to task difficulty.
OpenAI claims GPT-5.6 Sol scores 72.7% on the DeepSWE benchmark against Claude Fable 5's 69.9%, at lower estimated API cost. That's OpenAI's own figure.
She suggests tracking outputs in three buckets: ready to use, needs correction, needs escalation.
Why it matters: It's a sales argument wearing a CFO's clothes, and it lands right after a competitor published its own benchmark table claiming something different. But the underlying point survives the marketing. If your team is counting seats and tokens instead of finished work, you don't actually know whether any of this is paying for itself.
AI TOOLS
Useful AI Tools
Base44: No-code AI app builder that turns your words into fully functional apps in minutes.*
Chatbase: Build and deploy custom AI support agents for your business, no coding needed.*
ElevenLabs: Generate realistic, studio-quality voiceovers from text in any language or voice.*
Flora: Node-based creative canvas built for professional designers, agencies, and creative teams.*
Kick: Automates bookkeeping and tax preparation for business owners using AI.*
Magica: All-in-one AI platform with 5000+ tools for text, images, videos, audio, and lots more.*
WisprFlow: Voice-to-text AI that turns your speech into polished writing, so you can just talk instead of type.*
*Sponsored Listing
COMMUNITY
Featured Reader
Every newsletter, we feature how a reader is using AI to work smarter, save time, or make life easier.
Today's reader is Fiona M. from Galway, Ireland:
"I manage a small hotel and the reviews were eating me alive. Not writing them, reading them. Three sites, maybe forty a month, and by the time I'd got through them I'd forgotten what the first one said. So now I dump the whole lot in at the end of the month and ask it to pull out what people actually keep mentioning, not the one-offs. Turns out it wasn't the rooms at all, it was the breakfast finishing at half nine. Moved it to ten. Next month the complaints just stopped. Twenty years I've been doing this and I needed a computer to tell me people like a lie-in."
Want to be in the next newsletter? Tell us how you use AI here.
AI Meme
Go Deeper
AI Tools: A curated list of the AI tools worth knowing.
Instagram: The AI page they pray you’ll never find.
Facebook: For the ones who never left Facebook.
