Good morning, AI enthusiasts. The consumer backlash over OpenAI's Pentagon deal was loud, but this weekend marked an internal escalation: the company's robotics hardware lead resigned on principle.

Caitlin Kalinowski's departure is the first senior-level exit tied directly to the deal, and her public statement citing concerns over surveillance and lethal autonomy hits a bit harder than any App Store chart.

In today’s AI rundown:

• OAI's robotics lead exits over Pentagon deal

• The Rundown Roundtable: Our AI use cases

• How to build an AI case study generator

• Claude digs up 22 Firefox security flaws in two weeks

• 4 new AI tools, community workflows, and more

LATEST DEVELOPMENTS

OPENAI

🚪 OAI's robotics lead exits over Pentagon deal

Image source: Lovart / @kalinowski007 on X

The Rundown: OpenAI robotics director Caitlin Kalinowski just announced her resignation over the company’s controversial Pentagon deal, calling it a rushed move that skipped the guardrails on AI surveillance and lethal autonomy.

The details:

• Kalinowski joined OAI from Meta's AR glasses team in November, spearheading the rebuild of its robotics division that had previously shut down in 2020.

• She called the decision "about principle, not people", saying the deal was pushed through "without the guardrails defined" on AI in warfare.

• Kalinowski marks the first public “resignation” over the Pentagon deal, though VP of Research Max Schwarzer also departed last week for Anthropic.

• The backlash has hit fast on the consumer end, with Claude climbing to No. 1 on the App Store and ChatGPT cancellations soaring.

Why it matters: Plenty of users have ditched ChatGPT and spoken out since the Pentagon deal dropped — but Kalinowski is the first senior OAI voice to walk over it on principle. OAI can weather angry tweets and App Store slides, but a big resignation letter that name-drops "lethal autonomy" and "surveillance" hits a bit differently.

TOGETHER WITH YOU.COM

📖 AI Agents are your new employees

The Rundown AI agents are entering the workforce, but you wouldn’t expect a new employee to know everything on day one, would you? AI agents need onboarding too—in the form of metadata.

In this eBook, you’ll learn:

• How metadata management drives AI success

• Common pitfalls

• The ROI of proper metadata management

Get the eBook.

THE RUNDOWN ROUNDTABLE

💡The Rundown Roundtable: Our AI use cases

The Rundown: The Rundown Roundtable is a weekly feature where we poll members of The Rundown staff about how we use AI in our work and daily lives.

Rishi, Growth: We’re always looking for A/B test ideas for our landing pages and new ways to improve conversion rates while creating a better user experience. Recently, I tried an interesting AI workflow after coming across a public CRO skill on Twitter.

I installed it in Claude Code and then fed Claude screenshots of our landing page along with behavioral data from Microsoft Clarity, including scroll depth, heatmaps, and which buttons people were clicking most.

Using the CRO scorecard framework and the Clarity data, Claude generated a detailed analysis of the page and recommended five A/B tests we should prioritize, along with the reasoning behind each one. The insights were genuinely useful, and we have already taken action by launching one of the recommended tests.

Adrian, Developer: I’ve basically had the exact same haircut since high school, so this year I finally decided to experiment a bit with Nano Banana 2 and see what a different look might feel like. I uploaded my own portraits and started merging them with different models’ hairstyles. After 10+ rounds of virtual makeovers, I found out that the hairstyle that suited me best was… my current one.

AI TRAINING

⚙️ Build an AI case study generator

The Rundown: In this guide, you will learn how to use Claude to turn those old project files, like client emails, metrics exports, or memos, into information-rich case studies you can use to win more proposals.

Step-by-step:

1. Write project wrap memos when a project ends. Give an overview of results and key KPIs. If you’re busy, even a quick voice memo is better than nothing

2. Create a “Case Study Generator” project in Claude’s PC app with instructions: "You are a case study writer for [industry]. Your job is to take raw data and turn it into a case study using the challenge → solution → results framework"

3. Upload your project memo and prompt: “Generate a case study from these docs. Lead with strongest results in the headline. Use the challenge → solution → results framework. If any section is missing data, flag it with [NEEDS INPUT]”

4. Claude should generate a PDF case study for you. We found that it did a great job turning the memo into a slide deck or social media carousel post

Pro tip: We can pack a lot more into our system instructions, including a style guide and example cases. You can grab the templates here.

PRESENTED BY IBM

💪 How to build a strong AI-ready data foundation

The Rundown: IBM explains how unified, secured, and governed data access can help organizations move promising AI pilots to reliable enterprise scale.

In this guide, you’ll:

• Understand barriers that prevent AI pilots from scaling

• Learn why unified access to structured and unstructured data matters

• Explore a framework for building AI-ready data foundations

Download the guide.

ANTHROPIC & FIREFOX

🔒 Claude digs up 22 Firefox security flaws in two weeks

Image source: Anthropic

The Rundown: Anthropic revealed that Claude Opus 4.6 spent two weeks tearing through Firefox's codebase alongside Mozilla's team, turning up 22 vulnerabilities (14 high-severity) — with patches already live for hundreds of millions of users.

The details:

• Claude took just 20 minutes to flag its first flaw, and racked up 50 more by the time Anthropic's team finished confirming its initial find was real.

• Anthropic filed 112 reports across ~6K files in total — 14 rated high-severity by Mozilla, accounting for nearly 20% of Firefox's most serious patches all year.

• Claude also tried writing exploits, but only managed two working attacks in hundreds of attempts — both needing Firefox’s sandbox removed to function.

Why it matters: Firefox isn’t some new app; it’s a deeply tested open-source project with decades of audits and bounty programs — making Claude’s quick findings even wilder. While Claude wasn’t as strong at weaponizing its own exploits, Anthropic said that gap won’t last… Meaning the window to lock down codebases feels pretty urgent.

QUICK HITS

🛠️ Trending AI Tools

• 🗣️ Unwrap Customer Intelligence - Connect your entire organization to the true voice of the customer with AI-driven insights from customer feedback*

• 🔒 Codex Security - OpenAI's security agent to scan repos and patch bugs

• ⚙️ autoresearch - Andrej Karpathy's tool for AI-driven LLM training

• 🎆 Uni-1 - Luma’s model that reasons and generates across text and images

*Sponsored Listing

📰 Everything else in AI today

Luma unveiled Uni-1, its first model that combines reasoning and image generation in one architecture — in a major shift from the video-focused startup's roots.

Anthropic rolled out scheduled tasks in Claude Code, letting the coding agent run prompts on a loop to monitor builds, check logs, and auto-file PRs on a set cadence.

Cluely CEO Roy Lee admitted to fabricating the startup's revenue figures in a 2025 interview, publicly retracting claims after the AI ‘cheating’ tool pivoted to meeting notes.

The WSJ shared more on AI's role in the Iran conflict, reporting that the Army's 18th Airborne matched its Iraq-era targeting with 20 people instead of 2,000 with the tech.

Andrej Karpathy released autoresearch, an open-source repo that lets AI agents autonomously run and iterate on LLM training experiments in a loop on a single GPU.

COMMUNITY

🤝 Community AI workflows

Every newsletter, we showcase how a reader is using AI to work smarter, save time, or make life easier.

Today’s workflow comes from reader Julie H. in Kyle, TX:

"I made a day planner agent connected to my calendar and task lists. Every morning, I start by asking, “How does my day look?”, and the agent pulls all my meetings, uses my tasks list to schedule projects in between meetings, gives me time for deep focus and email catch-up, and even makes sure to schedule a lunch break.

It flags things I may have missed the day before. If I have big gaps in the day, it looks ahead and suggests items I can get ahead on, like tasks due the next day or meetings needing prep time. I can give feedback on things I want to add, move, or remove, and it adjusts until I have a solid plan for my day. If I move things in my schedule, it will also update my calendar and task list for me so everything stays aligned."

How do you use AI? Tell us here.

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: OpenAI’s ‘best model ever’ goes live

• Read our last Tech newsletter: Meta sued over Ray-Ban privacy

• Read our last Robotics newsletter: Waymo has a big problem in Austin

• Today’s AI tool guide: Build an AI case study generator

• RSVP to our next workshop Thursday @ 2PM EST: AI coding bootcamp pt. 1

See you soon,

Rowan, Joey, Zach, Shubham, and Jennifer — the humans behind The Rundown

Good morning, AI enthusiasts. A week ago, Spotify’s co-CEO claimed their best devs haven’t written a “single line of code” this year — echoing a wave of execs that describe AI coding agents as the future of software. 

The shift is happening – there’s no doubt about that. But bringing AI into real engineering workflows is more nuanced than hitting a switch and going on autopilot.

To better understand what’s changing (and what isn’t), we sat down with Rajeev Rajan, CTO of Atlassian, the company behind popular collaboration tools such as Jira, Confluence, and Loom. We got insights on their recent software development agent Rovo Dev and discussed how human roles evolve in an AI-native world.

In today’s AI rundown:

• The non-negotiable cost of coding with AI

• When AI writes code, what will an engineer do?

• Building agentic AI for developer joy

• What happens when AI makes a mistake?

• Truth about the “death of SaaS” theory

• Quick hits with Rajeev

LATEST DEVELOPMENTS

WORKFLOW REDESIGN

🤖 The non-negotiable cost of coding with AI

The Rundown: As more enterprises adopt coding agents, Rajan says teams will have to redesign their workflows with both human ownership and safety systems, so that AI moves faster without sacrificing consistent, production-grade quality.

Cheung: AI is generating more code than ever, but research shows 45% of it still contains security flaws. How exactly should teams leverage AI coding agents without trading quality for speed?

Rajan: There’s an undeniable trend toward an AI‑native software development lifecycle. But if you let quality slide, you’re just moving faster toward incidents and customer pain. It’s less about “AI vs. quality” and more about “how do we redesign the workflow on the assumption that AI is in the loop by default?”

In code review, we are entrusting AI to catch bugs, enforce coding standards, and explain complex changes. For example, Rovo Dev helped reduce PR cycle time by 45% and auto-resolved 51% of potential security vulnerabilities. The nature of review is changing here: instead of humans reading every line of a peer’s code, it’s about a human owner reviewing an agent’s work.

Rajan added: At deployment, if AI is helping you generate and ship more code, your safety systems have to keep pace. Think: smaller batches, heavier CI, stronger observability, and fast rollbacks. You can’t operate in a black‑box scenario.

Why it matters: Speed only becomes an advantage when the system behind it is airtight. The right approach with AI coding agents is to treat them as core enterprise infrastructure — which means designing workflows around them, building in safety from day one, and holding output to the same standards as human-written code.

ROLE SHIFT

❓️ When AI writes the code, what will an engineer do?

Image: Kiki Wu / The Rundown

The Rundown: With writing code no longer the bottleneck, Rajan believes that the next big opportunity for engineers is stepping into more strategic functions (like planning to execution) and designing better systems with AI in the loop.

Cheung: How much of code will be AI-written by 2028, and what does the role of a software engineer look like with that change?

Rajan: By 2028, I would not be surprised if most new code in large companies is AI-generated. I say that as someone who fell in love with writing code early in my career and still remembers the joy of seeing something work for the first time.

We’re seeing a shift where every engineer is a tech lead, orchestrating systems and agents. Engineers now spend more time driving clarity and owning what happens “left of code” and “right of code” – from planning and design on one side to testing, rollout safety, and operations on the other.

Cheung: What about new grads entering the field — does AI help them or hurt them?

Rajan: Focusing on the right fundamentals and adopting the AI-native way of working will give new grads a big advantage — potentially allowing them to leapfrog senior developers who haven’t adopted AI ways of working yet. Your edge will come from judgment: knowing when to trust the AI and when to challenge it.

Why it matters: As AI writes more code, the moat for engineers moves from actually typing to instead framing problems, designing systems, and maintaining oversight. Rajan’s leapfrog point is especially interesting: the advantage may not go to the most senior person, but to whoever learns to orchestrate AI fastest.

DEVELOPER JOY

⚙️ Building agentic AI for developer joy

The Rundown: Atlassian kicked off its internal journey to improve “Developer Joy,” raising developer satisfaction scores from 49% to 83%. With teams moving faster and feeling more empowered to make changes, Rajan shared how this renewed sense of ownership led to direct product improvements with Rovo Dev.

Cheung: Why did you decide to focus on developer joy, and how did you actually measure it?

Rajan: When I joined Atlassian, we chose to frame developer productivity as ‘Developer Joy’. If developers are frustrated, blocked, or taken out of their flow, it doesn’t matter what productivity metric you pick — you’re not going to get great outcomes.

We track this with regular satisfaction surveys and hard metrics tied to pain points. Developer satisfaction has gone from 49% to 83%, and we see that show up in the work. For example, by focusing on one of the biggest friction points, the Confluence backend team cut its full server build time by more than 60%. We see these kinds of investments as core to our ability to ship value to customers.

Cheung: When you tested early versions of Rovo Dev internally, what did engineers push back on?

Rajan: Early on, the feedback we got on Rovo Dev was that parts of the experience felt like ‘magic’ in the wrong way. It would do something useful, but you could not see enough of the work it did to get there.

We actually scrapped and reworked an early ‘one click, do it all’ flow because our own teams would not touch it without more transparency and control. They wanted a way to understand each step the agent took, how their instructions led to different outcomes, and the ability to steer the agent. That pushed us hard toward agent sessions you can inspect and experiences that keep developers in the loop.

Why it matters: Rajan’s early Rovo Dev story highlights how critical internal feedback loops are when adopting agentic AI. The more teams listen to users (and iterate on what feels opaque, risky, or frustrating), the stronger and more trustworthy the system becomes. Iteration goes hand-in-hand with developer trust in an AI-native world.

ACCOUNTABILITY

🧠 What happens when AI makes a mistake?

The Rundown: Rajan says powerful AI agents should be deployed to production only when there’s clear human ownership and a way to track, monitor, and steer the system’s behavior, creating an accountability layer that moves as fast as the AI itself.

Cheung: As AI agents take on more autonomous work, how do you maintain accountability, and why can’t “the AI did it” ever be an excuse?

Rajan: When something goes wrong, ‘the AI did it’ can’t be the answer, because AI doesn’t own customer trust – we do.

As we bring autonomous AI into our workflows, we have to be explicit about accountability: every AI-assisted decision or action has to have a clear human owner. If we can’t understand or observe how an AI is behaving, it doesn’t belong in a critical path.

We put guardrails and observability around AI, we log and audit its actions, and we make sure teams treat it like any other powerful tool: you understand failure modes, you monitor it, and you don’t ship without ownership. AI can help move faster, but it doesn’t replace judgment and responsibility.

Why it matters: With something as powerful as agentic AI, accountability is non-negotiable. If you get it right, you unlock speed, trust, and durable customer confidence. But if you get it wrong, the consequences can be massive — because autonomous systems can amplify mistakes just as quickly as they amplify progress.

SAAS STORY

⚡️ Truth about the “death of SaaS” theory

The Rundown: Despite the rise of AI coding agents, Rajan argues SaaS tools aren’t going anywhere. In fact, he believes these tools will get stronger — with AI working across projects and controls by tapping contextual insights.

Cheung: What’s your take on the whole “saaspocalypse” narrative that AI agents will kill SaaS altogether?

Rajan: The idea that one person in your team can vibe code an in-house solution over a weekend and replace a mature SaaS solution you’re paying for, in my view, is overrated.

When customers buy software-as-a-service, they are not just buying the code; they buy workflows, shared context, security, compliance, and reliability. That is where well-designed SaaS products still matter a lot. What AI actually does is make great SaaS even more valuable.

Rajan added: Your projects, docs, tickets, and conversations live in those systems, and AI can now move across them, automate the boring parts, and orchestrate agents around trusted workflows and controls. So I am more interested in SaaS that becomes AI-native than in hot takes that SaaS is dead.

Why it matters: The debate is still on, but one thing’s clear: as AI agents grow more capable, the foundation of SaaS will become even more important. AI will evolve the platforms that already hold an organization’s workflows and institutional knowledge, serving as trusted systems of record.

LIGHTNING ROUND

⚡️ Quick hits with Rajeev

Most underrated AI trend right now?

Rajan: Lots of AI products today are designed for a single-player system. We see much greater potential in how AI helps entire teams work better together — allowing important context to flow across a team of humans and agents.

Something you believe about AI that most people in tech would disagree with?

Rajan: I think AI will make engineering more human, not less. A lot of people worry we will lose the craft — I believe we will spend less energy on repetitive implementation and more time on strategic, creative work and collaboration.

Advice for teams struggling with developer burnout?

Rajan: Start by fixing one concrete, high-friction problem that impacts your team. You’d be surprised by how quickly chipping away at issues like slow build times and noisy tooling can multiply and have a greater impact.

You worked at Microsoft for 20+ years, then led engineering at Meta. What did each teach you about building great teams?

Rajan: Microsoft taught me the value of deep technical rigor and building platforms that stand the test of time. Meta taught me how powerful it is when you pair strong engineering talent with a bias for fast iteration and learning. At Atlassian, I try to combine both: long-term architecture with a culture that ships, learns, and adapts quickly.

See you soon,

Rowan, Joey, Zach, Shubham, and Jennifer — the humans behind The Rundown

Good morning, tech enthusiasts. Meta’s Ray-Ban smart glasses are at the center of a nightmarish privacy lawsuit: a proposed U.S. class action alleges the company let human reviewers access users’ intimate recordings, despite marketing that suggested stronger protections.

The case is raising a simple, unsettling question for smart-glasses users: when the camera is on, who might really be watching?

In today’s tech rundown:

• Meta’s AI glasses hit with privacy suit

• Apple’s $599 MacBook for Chromebook crowd

• Oura acquires gesture-recognition startup

• Science Corp nabs $230M for brain implant

• Quick hits on other tech news

LATEST DEVELOPMENTS

META

👓 Meta’s AI glasses hit with privacy suit

Image source: Ideogram / The Rundown

The Rundown: Meta is facing a proposed U.S. class-action lawsuit alleging that its Ray-Ban Meta AI smart glasses secretly routed intimate user videos to low-paid human reviewers, violating the company’s own privacy promises.

The details:

• A proposed U.S. class-action lawsuit accuses Meta and Luxottica of misleading buyers about the privacy protections of Ray-Ban smart glasses.

• Plaintiffs say Meta failed to disclose that contractors and employees could view user recordings, including nudity, sex acts, and people using the bathroom.

• The case follows reports that low-paid reviewers in Kenya were required to watch and label intimate clips as part of Meta’s safety and AI-training pipeline.

• Regulators, including authorities in the U.K., are also probing whether Meta’s data practices for the glasses meet transparency and consent standards.

Why it matters: The complaint says Meta misled customers about this human-review pipeline and put millions at risk of stalking, extortion, and identity theft, while Meta counters that contractors review shared Meta AI content to improve the glasses’ user experience and that this human involvement is disclosed in its policy fine print.

APPLE

🍏 Apple’s $599 MacBook for the Chromebook crowd

Image source: Apple

The Rundown: Apple launched a $599, iPhone-chip-powered MacBook Neo that brings bright colors and bargain pricing to directly challenge low-cost Windows laptops and Chromebooks. It’s set to hit stores on March 11.

The details:

• The Neo uses an iPhone‑class A‑series chip paired with up to 16GB of unified memory and fast NVMe storage, aiming for all‑day battery life.

• Apple is offering the Neo in multiple bright colors, visually positioning it closer to iPads and older plastic MacBooks than its current pro-focused lineup.

• The machine targets students and first-time buyers who currently gravitate toward Chromebooks and cheaper Windows laptops.

• By undercutting traditional MacBook prices, Apple is moving aggressively into a segment long dominated by PC OEMs and Google’s education partners.

Why it matters: Apple is putting real pressure on bargain Windows laptops and Chromebooks with a $599 MacBook that still feels like a “real” Mac. By pairing an iPhone-class chip with a colorful, fanless design and long battery life, the MacBook Neo aims to hook first-time buyers without undercutting its pricier line.

OURA

👊🏼 Oura acquires gesture-recognition startup

Image source: Oura

The Rundown: Smart ring maker Oura just acquired Helsinki-based gesture-recognition startup Doublepoint to fuse voice and micro‑hand movements into its rings, turning the health tracker into a discreet control hub for next‑gen wearable AI.

The details:

• The deal is designed to layer gesture recognition and voice controls on top of Oura’s continuous health sensing to trigger features without screens.

• All of Doublepoint’s team, including its four co-founders, will join Oura, forming a new hub for interaction design inside the Finnish wearable company.

• Oura, valued at $11B, has sold more than 5.5M rings and expects sales to top $1.5B in 2026 as smart ring shipments jumped about 51% last year.

• This is Oura’s fourth recent acquisition, following Sparta Science, Veri, and Proxy, each adding a new layer to what was primarily a sleep tracker.

Why it matters: Oura CEO Tom Hale said gesture controls, paired with voice inputs, could become central to future versions of the ring, though meaningful implementation isn’t happening right away. Still, the acquisition helps Oura defend its smart ring lead in the category, as wearables start moving toward a post-screen interaction model.

SCIENCE CORPORATION

🧠 Science Corp nabs $230M for brain implant

Image source: Science Corp (PRIMA retinal implant)

The Rundown: Neuralink co-founder Max Hodak’s neurotech startup Science Corporation closed a $230M Series C, valuing the company at roughly $1.5B and putting serious fuel behind its bid to restore sight to patients blinded by retinal disease.

The details:

• The funds will accelerate clinical trials and regulatory approvals in the U.S. and Europe for PRIMA, a subretinal chip paired with camera-equipped glasses.

• In early trials with 47 patients, PRIMA has helped people with advanced macular degeneration recognize letters, numbers, and short words.

• Science Corp. is seeking a CE mark in the EU, aiming for mid‑2026 approval with Germany as its first launch market, while it continues talks with the FDA.

Why it matters: Science Corporation’s side projects — from cortical interfaces to organ-preservation tech — show it’s thinking beyond implants to a neurotech platform. That puts it in more direct competition with Neuralink and Precision Neuroscience, which are also racing to own the hardware layer between brains and machines.

QUICK HITS

📰 Everything else in tech today

Oracle is reportedly preparing to cut thousands of jobs as it grapples with a cash crunch caused by massive spending on AI data centers and related infrastructure.

A global Amazon outage left tens of thousands of shoppers facing failed payments and checkout glitches, with Downdetector logging around 20K problem reports.

Online retailer Quince is reportedly in talks with investors to raise a new funding round that would more than double its valuation to above $10B.

Microsoft confirmed that its next-gen Xbox, codenamed Project Helix, is a high-performance console that will run both Xbox and PC games.

U.S. defense tech firm Anduril is raising a multibillion-dollar round that could double last year’s $30B valuation to about $60B. 

Researchers built a stable, ultra-bright 300-nanometer OLED pixel using a nano-antenna, enabling full HD displays small enough to fit on a grain of sand.

London-based smartphone maker Nothing launched new phones and headphones in bright colors to counter what its founder calls a “boring” tech market.

Whoop, the maker of connected fitness bands, plans to increase its workforce by about 75% in 2026 by hiring more than 600 people to fuel growth ahead of a likely IPO.

Cluely co-founder and CEO Roy Lee admitted on X that the $7M annual recurring revenue figure he gave TechCrunch last summer was a deliberate lie.

BYD unveiled its Blade Battery 2.0 pack, which can fully charge in just about 10 minutes but only when paired with its new 1.5-megawatt Flash Charging stations.

COMMUNITY

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: OpenAI's 'best model ever' goes live

• Read our last Tech newsletter: Apple’s budget iPhone goes AI first

• Read our last Robotics newsletter: Waymo has a big problem in Austin

• Today’s AI tool guide: Turn an investment memo into a polished slide deck

• RSVP to our next workshop on March 12: AI coding bootcamp pt. 1

See you soon,

Rowan, Joey, Zach, Shubham, and Jennifer — The Rundown’s editorial team

Good morning, AI enthusiasts. OpenAI needed a win after a rough stretch of public sentiment — and GPT-5.4 looks like a big one.

The company's new top model outperforms humans on desktop tasks and across 83% of job-specific evaluations, with new highs across math, science, coding, and reasoning to match. OAI researcher Noam Brown's take: "We see no wall."

In today’s AI rundown:

• GPT-5.4 beats humans at their own desktops

• Netflix acquires Ben Affleck's AI filmmaking startup

• Turn an investment memo into a polished slide deck

• Anthropic’s early-warning system for AI job loss

• 4 new AI tools, community workflows, and more

LATEST DEVELOPMENTS

OPENAI

🧠 GPT-5.4 beats humans at their own desktops

Image source: OpenAI

The Rundown: OpenAI just rolled out GPT-5.4, the company’s new top model with major upgrades on desktop tasks, coding, reasoning, science, math, and more — with VP of Science Kevin Weil calling it "our best model ever".

The details:

• OAI shipped GPT-5.4 just two days after rolling out 5.3 Instant as the default chat model, available now as GPT-5.4 Thinking for Plus, Team, and Pro users.

• The model scored 75% on OSWorld-V, which tests real desktop navigation — 3 points above the human baseline of 72.4% and 2x of what GPT-5.2 managed.

• 5.4 also supports up to 1M tokens of context and a new x-high reasoning effort setting, letting agents plan and execute across longer tasks that take hours.

• GPT-5.4 won or matched against professionals 83% of the time on GDPval, a knowledge-work benchmark across 44 jobs — up from 71% for GPT-5.2.

Why it matters: OpenAI needed a win after a rough week of sentiment, and GPT 5.4 looks like one — with performance that looks to take the next step up the frontier, particularly for desktop use cases to push forward agentic abilities. The launch also comes with a big statement from OAI researcher Noam Brown: "We see no wall”.

TOGETHER WITH AIRIA

💼 Redefine enterprise AI with Airia

The Rundown: Airia delivers unified AI security, orchestration, and governance built for enterprises, accelerating AI adoption. Deploy agents fast while maintaining control, bridging the gap between innovation and compliance.

Dive deeper with Airia’s 2026 State of AI Report, which reveals:

• Key insights on enterprise AI adoption trends

• Critical challenges and emerging opportunities ahead

• A comprehensive guide for navigating AI transformation

Learn more.

NETFLIX & BEN AFFLECK

🎬 Netflix acquires Ben Affleck's AI filmmaking startup

Image source: Netflix

The Rundown: Netflix just acquired InterPositive, a stealth AI filmmaking company Ben Affleck started in 2022 — bringing all 16 staffers and Affleck himself aboard as senior adviser in a rare acquisition for a streaming giant.

The details:

• InterPositive's tech trains models on a production's own footage, then handles post work like relighting, swapping backgrounds, and fixing continuity errors.

• Affleck said he was shocked by how much engineering talent was pouring into AI video, "but no artistic, no filmmaking information whatsoever”.

• The actor emphasized that his company’s tech is “not generating video from nothing”, instead learning from the existing filmed shots and actors.

• Affleck appeared on the JRE Podcast last month, saying he "can't stand" what AI writes, and that the tech would be more of a tool for production workflows.

Why it matters: Hollywood has spent the AI boom either hiding the tech’s use or railing against it, but an Oscar-winning industry leader putting his reputation on a tool could go a long way to shifting sentiment. For all the “AI killed Hollywood” X posts, the real upgrades come from the production workflow aspects Affleck is addressing.

AI TRAINING

📋 Turn an investment memo into a polished slide deck

The Rundown: In this guide, you will learn how to use Manus to turn a structured investment memo into a professional, ready-to-present slide deck in Google Slides. You can try this with any type of memo with structured, formulaic sections.

Step-by-step:

1. Start with a memo PDF with sections like: market analysis, team background, business diligence, traction, SWOT analysis, and deal dynamics. We used Sequoia’s 2014 DoorDash memo for our test

2. Go to manus.im and sign up, then click New task > Create slides, attach your PDF, and prompt: “Turn this investment memo into a slide deck”

3. That’s it. Manus will create the deck for you, and it usually will one-shot the layout. You can edit or present the deck in Manus, or ask it to update it for you

4. When done, you can download your slides as PPTX, PDF, or click “Convert to Google Slides,” and it will save the deck perfectly in Google Slides for you

Pro tip: Attach an example deck with your memo and tell Manus to style your new deck in the same way.

PRESENTED BY BLAND AI

☎️ How Soulja Boy automated his voice with AI

The Rundown: Soulja Boy just became the first rapper to automate his voice with AI. Bland AI, a voice AI company out of San Francisco, bought his voice to let anyone call him — generating 30M+ views in 24 hours and thousands of enterprise signups.

Even if you don't want Soulja answering your business calls, Bland allows you to:

• Clone your best reps' voice to scale their success across your entire business

• Deploy AI voice agents that sound human, without adding headcount

• Get started for free on a platform trusted by thousands of enterprises

Call Soulja Boy to try it out at (415) 480-0000 or sign up free at bland.ai.

AI RESEARCH

📊 Anthropic’s early-warning system for AI job loss

Image source: Anthropic

The Rundown: Anthropic published a study on AI’s job impact, cross-referencing what AI could automate against what people are using Claude for — finding that while mass layoffs haven't hit, the youngest workers in are already getting squeezed out.

The details:

• The study uses "observed exposure," a metric that gauges AI job displacement by comparing tasks AI can do to what they are already automating.

• Computer programmers top the exposure list at 75% task coverage, followed by customer service reps and data entry workers at 67%.

• Roughly a third of the U.S. workforce sits at zero AI exposure right now, largely in hands-on roles like cooks, bartenders, and lifeguards.

• No broad unemployment spike has appeared since ChatGPT's launch in 2022, but hiring into exposed fields for 22-to-25-year-olds fell 14% in that time.

Why it matters: Anthropic CEO Dario Amodei has not been subtle about what he believes is looming on the jobs front due to AI, and we’ve already seen several industry stock prices tank this year following Claude releases. But even with the warnings, the world still feels drastically underprepared for the disruption coming.

QUICK HITS

🛠️ Trending AI Tools

• 🗣️ Unwrap Customer Intelligence - Connect your entire organization to the true voice of the customer with AI-driven insights from customer feedback*

• 🧠 GPT-5.4 - OpenAI's flagship reasoner with native computer use, 1M context

• 🎞️ Luma Agents - Luma's AI agent for scaling creative output across teams

• 🎥 LTX-2.3 - Lightricks' video AI upgrade with more detail, cleaner audio

*Sponsored Listing

📰 Everything else in AI today

The Pentagon officially labeled Anthropic ‘supply chain risk’, which the company plans to challenge in court, coming amid reports that both sides had resumed deal talks.

Lightricks released LTX-2.3, an upgrade to its powerful open-source video model, and LTX Desktop, a free local video editor built on the same engine.

Google released an open-source CLI for its full Workspace suite, with 40+ built-in agent skills designed for easy integration into agentic platforms.

Ex-OpenAI chief research officer Bob McGrew is raising $70M at a $700M valuation for Arda, his startup building an AI platform to automate factory floors with robots.

Meta is being sued after an investigation found that overseas contractors reviewing Ray-Ban AI smart glasses footage were seeing nudity and other private user content.

COMMUNITY

🤝 Community AI workflows

Every newsletter, we showcase how a reader is using AI to work smarter, save time, or make life easier.

Today’s workflow comes from reader Tineke B. in Auckland, NZ:

"I just created a website with AI that's clocked up 5,000 organic visits in less than a week: To find out ferry timetables in Auckland, you had to visit the Auckland Transport website, click numerous links, and then download a PDF (yes, a PDF!) to view the ferry timetables.

This was super annoying for me, so I used Claude to create a site for it. I shared it on a few local FB groups, and it's taken off. Over 5,000 visits in less than a week and 100s of comments on social media thanking me for making people's lives easier. It only took me a couple of hours and has solved a real (first-world) problem for commuters.”

How do you use AI? Tell us here.

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: Amodei torches OpenAI in leaked memo

• Read our last Tech newsletter: Apple’s budget iPhone goes AI first

• Read our last Robotics newsletter: Waymo has a big problem in Austin

• Today’s AI tool guide: Turn an investment memo into a polished slide deck

• RSVP to our next workshop on March 12: AI coding bootcamp pt. 1

See you soon,

Rowan, Joey, Zach, Shubham, and Jennifer—the humans behind The Rundown

Good morning, robotics enthusiasts. Waymo has long been the gold standard of autonomous driving. But federal investigators are circling after a robotaxi rolled past a school bus of boarding children — and just this Sunday, another blocked an ambulance racing to an Austin mass shooting.

The code will be updated, but can the company update confidence, too?

In today’s robotics rundown:

• Waymo has a school bus-sized problem

• Tether-backed Neura chases billion-dollar raise

• Noble exits stealth with a heavy-lifting humanoid

• Scientists built a farm bot that rots into the soil

• Quick hits on other robotics news

LATEST DEVELOPMENTS

WAYMO

🚌 Waymo has a school bus-sized problem

Image source: Ideogram / The Rundown

The Rundown: Waymo’s driverless taxis in Austin are under intense scrutiny after one failed to stop for a school bus and another blocked an ambulance rushing to a mass‑shooting scene, raising fresh questions about how these vehicles interpret risk.

The details:

• In January, a robotaxi illegally passed a stopped school bus while children were boarding, after a remote human agent mistakenly cleared it to proceed.

• Austin ISD cameras have caught roughly two dozen similar Waymo school-bus violations since September.

• NHTSA and NTSB have both opened formal investigations, with the January incident now the centerpiece of an active federal probe.

• In a separate episode, a driverless Waymo briefly blocked an ambulance, forcing a police officer to commandeer the vehicle and move it out of the way.

Why it matters: Waymo has filed a software recall, insists it is improving performance around school buses and emergency scenes, and has declined detailed public comment on the ambulance incident even as local officials and safety advocates question whether its “safety‑first” AV stack is ready for crowded urban chaos.

NEURA ROBOTICS

🤖 Tether-backed Neura chases billion-dollar raise 

Image source: Neura Robotics

The Rundown: German humanoid startup Neura Robotics is reportedly closing in on one of Europe's largest-ever robotics financings — a €1B (~$1.1B) round backed by stablecoin giant Tether.

The details:

• The deal would value the Metzingen-based company at roughly €4B ($4.3B), making it one of Europe’s highest valued robotics players.

• Neura has already locked in an order book of around €1B from industrial customers for its transport robots, cobots, and robotic arms.

• The company's edge: tight in-house integration of sensors, control electronics, and AI software into what it calls “cognitive robots.”

• The raise follows a $130M Series B last year that pushed total funding past $200M and helped Neura double its headcount to over 300.

Why it matters: A Tether-led round would mark one of the crypto firm’s boldest moves outside digital assets, following earlier bets on Blackrock Neurotech and Italian humanoid startup Generative Bionics. For Neura, a 10-figure raise would put a European humanoid maker in the same fundraising weight class as top U.S. rivals.

NOBLE MACHINES

💪 Noble exits stealth with a heavy-lifting humanoid

Image source: Noble Machines

The Rundown: Ex‑SpaceX and Apple veterans’ startup Noble Machines emerged from stealth this week with its no‑frills, heavy‑lifting Moby humanoid already working real shifts on factory floors just 18 months after founding.

The details:

• Noble Machines — formerly Under Control Robotics — says its robots are already deployed at an undisclosed Fortune Global 500 manufacturer.

• Founded by engineers from SpaceX, Apple, NASA, and Caltech, the Sunnyvale-based startup is focused on real factory and construction environments.

• Its Moby humanoid is built for hazardous, physically demanding work across manufacturing, construction, logistics, energy, and semiconductor plants.

• Running end-to-end autonomy on a single NVIDIA Jetson Orin with an Isaac-based training loop, the robot uses whole-body control to learn new tasks.

Why it matters: Noble Machines says its Moby bot uses whole-body AI control and language-based learning to quickly pick up new jobs while hauling 60 lbs. across rough terrain, putting it in the same lifting league as Boston Dynamics’ Atlas and ahead of Agility’s Digit and Figure 3 — and a next-gen Moby is already on the way.

ROBOTICS RESEARCH

🌾 Scientists built a farm bot that rots into the soil

Image source: Ideogram / The Rundown

The Rundown: Scientists just built a fully compostable soft robot designed for real-world farm use — one that monitors crops for a full growing season, then quietly dissolves into the soil it just analyzed rather than becoming e-waste.

The details:

• The system uses biodegradable electronics and soft materials that survive real farm conditions to monitor plant health and the environment.

• Robotic fingers or soft grippers are designed for in-field plant monitoring tasks like tracking moisture, nutrients, or pollutants at high spatial resolution.

• When its job is done, it breaks down into benign byproducts that can act as soil nutrients, turning the sensor itself into part of the system it was monitoring.

• The work pushes soft robotics toward “ecoresorbable” designs, where robots are treated as temporary infrastructure rather than permanent hardware.

Why it matters: Digital agriculture has long faced a tradeoff: durable sensors create e-waste, biodegradable ones can’t hack real soil conditions. This bot looks to have cracked that problem. Fully compostable components are now field-ready, moving degradable soft robotics out of the lab and into actual dirt.

QUICK HITS

📰 Everything else in robotics today

Amazon cut at least 100 white-collar jobs in its robotics division, separate from the 16K corporate layoffs announced in January.

Hyundai unveiled a firefighting robot designed to navigate dangerous environments, withstand intense heat and smoke, and support human crews.

Chinese autonomous driving startup Momenta confidentially filed for a Hong Kong IPO, shifting its long-planned listing away from New York.

WeRide, Baidu’s Apollo Go, and Pony.ai paused robotaxi services in parts of the Middle East, pulling back from Dubai amid regional security tensions.

Next-gen Coco Robotics delivery bots are rolling out in LA, with larger, tougher, more autonomous robots designed to better handle the wear and tear of city streets.

Qualcomm CEO Cristiano Amon said robotics, driven by its new Dragonwing AI chips, will become a major business for the company within two years.

Singapore-based startup Asimov opened preorders for its open-source Asimov humanoid, a roughly 1.2-meter-tall, 25‑plus‑DOF Unitree G1 rival.

Pony.ai claimed its seventh‑generation robotaxi has reached per‑vehicle profitability in Shenzhen, its second Chinese tier‑one city after Guangzhou.

Google DeepMind launched its first accelerator in Europe, a three‑month, equity‑free program for early‑stage European robotics startups.

China built a massive high‑precision 3D face database plus an AI model to give humanoids more natural, expressive faces.

The UK Atomic Energy Authority and CERN built a mouse-sized robot, PipeINEER, to crawl through LHC pipes and use AI vision to spot faults quickly and safely.

China flight-tested an 8-foot-long experimental drone built largely from bamboo, aiming to create a cheap, lightweight, and quickly deployable platform.

COMMUNITY

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: Amodei torches OpenAI in leaked memo

• Read our last Tech newsletter: Apple’s budget iPhone goes AI-first

• Read our last Robotics newsletter: Bio-robotic spy roaches

• Today’s AI tool guide: Turn any CSV into an Excel dashboard

• RSVP to our next workshop on March 12: AI coding bootcamp pt. 1

See you soon,

Rowan, Joey, Zach, Shubham, and Jennifer — The Rundown’s editorial team

Good morning, AI enthusiasts. It took less than a week for OpenAI’s Pentagon deal to go from announced to rewritten, and Sam Altman is calling the fallout “really painful.”

With protests outside the company’s HQ, mass cancellations, and its detailed contract revisions bringing as many questions as answers, the rushed agreement is turning what was supposed to be a power move into OpenAI’s biggest brand crisis in years.

In today’s AI rundown:

• OpenAI walks back Pentagon details after backlash

• ChatGPT upgrade fixes the ‘cringe’ problem

• Create killer thumbnails with Midjourney

• Google’s new 3.1 Flash-Lite pairs speed, cost, intelligence

• 4 new AI tools, community workflows, and more

LATEST DEVELOPMENTS

OPENAI

🪖 OpenAI walks back Pentagon details after backlash

Image source: Sam Altman on X

The Rundown: OpenAI CEO Sam Altman just posted a note on X detailing significant revisions to the company’s initial Pentagon contract, coming amid employee pushback, user cancellations, and a surge of sign-ups to Anthropic following the deal.

The details:

• OAI’s original agreement used the same language Anthropic had refused, finalized within 24 hours following the Pentagon’s ban of its rival.

• Altman called the deal rushed and said it looked “opportunistic and sloppy,” adding that he would “rather go to jail” than follow an unconstitutional order.

• Research scientist Noam Brown clarified that OAI “will not be deploying to the NSA or other DoW intelligence agencies for now,” as loopholes are addressed.

• Altman held an all-hands on Tuesday, calling the deal “complex but the right decision with extremely difficult brand consequences and negative PR for us.”

Why it matters: It’s been quite the headache for OAI post-Pentagon deal, with Altman admitting the company moved too fast — and consequences ranging from bad optics to massive consumer backlash and protests outside its SF offices. The amended contract language is a start, but the brand damage feels like it’s already been done.

TOGETHER WITH VISA

🤝 Get to know Visa’s Trusted Agent Protocol

The Rundown: The trust layer for agentic commerce. Visa’s Trusted Agent Protocol allows AI agents to verify who they are, who they represent, and what they’re authorized to do — so merchants can interact with agents with confidence, and humans stay in control.

Visa’s Trusted Agent Protocol helps deliver: 

• Cryptographic verification of AI agent identity and intent

• Clear authorization and accountability for agent‑initiated actions

• Trust signals to help merchants distinguish legitimate agents from bad actors

Learn more here.

OPENAI

🚀 OpenAI’s ChatGPT upgrade fixes the ‘cringe’ problem

Image source: OpenAI

The Rundown: OpenAI just released GPT-5.3 Instant, overhauling its default ChatGPT model to all users with a new release that prioritizes how the AI talks over how it thinks — tackling the preachy personality that’s frustrated users for months.

The details:

• OAI built 5.3 Instant around conversational quality, dialing back unnecessary refusals and what the company itself called a ‘cringe’ tone in responses.

• The update also claims a reduction in hallucinations, with rates down by over 25% on web search and nearly 20% on internal knowledge on benchmarks.

• OAI says 5.3 Instant is also a “stronger writing partner,” additionally providing better web answers and presentation of information.

• The company also teased a more significant model jump coming soon, posting an easter egg on X saying “5.4 sooner than you Think.”

Why it matters: Anthropic’s Super Bowl ads were controversial, but their poking at ChatGPT’s personality may have cut a bit too close to the truth — with OAI’s new rollout prioritizing moving its personality away from the “cringe” tone. But with uninstalls up 295% after the Pentagon drama, personality may not be the most pressing issue.

AI TRAINING

🎨 Create killer thumbnails with Midjourney

The Rundown: Learn a quick prompting and editing system to make thumbnails that stand out, but don’t look like AI slop. Whether it’s for your YouTube channel, a blog, or even your social feeds, learning to quickly make good thumbnails pays dividends.

Step-by-step:

1. Open Claude or any chat agent. Tell it your video topic and ask it to generate four Midjourney prompts with this structure: [Person + Expression] +[Action/Prop] + [Setting] + [Lighting] + [Composition with negative space] --ar 16:9.

2. Paste each prompt into the Midjourney web app. When you see a generation you like, click on it and try selecting vary > subtle or vary > strong until you get the perfect variation. Save the image to your desktop.

3. Key step: the design. Remember, simplicity is essential to this technique. Add 1–5 words of big, click-worthy text in Canva.

4. Bring the opacity down 5% and use a blocky font. Go with white or yellow text and optionally a black outline.

Pro tip: If you can’t get the text to fit right on your image, tell Claude to come up with a new prompt that will create that empty space for you.

PRESENTED BY GLEAN

⚙️ From AI ambition to secure agents at scale

The Rundown: Join Glean’s virtual Security Showcase on March 12 to close the gap between AI ambitions and security readiness. Get a practical blueprint for governing AI agents at scale from the leaders building, securing, and deploying them in production.

Register for Glean’s Security Showcase to:

• Hear Cvent’s CIO and CISO in a fireside chat on real-world AI agent risk decisions.

• Get access to a new security framework, designed for governed AI agents at scale.

• See what’s new in agent controls, data protection, and private‑by‑design deployment for enterprise AI.

Register now.

GOOGLE

🔦 Google’s new 3.1 Flash-Lite pairs speed, cost, intelligence

Image source: Google

The Rundown: Google just rolled out Gemini 3.1 Flash-Lite, the company’s fastest entry in its Gemini 3 lineup that provides a near-instantaneous feel and upgraded intelligence while undercutting rivals on price.

The details:

• Flash Lite rounds out Google’s tiered Gemini 3 release weeks after Pro, giving a budget option for high-volume work that doesn’t need a flagship model.

• Lite scored a 12-point jump on the Artificial Analysis Intelligence Index over its predecessor, beating even larger prior-gen Gemini models on reasoning.

• The model costs 1/4 of Anthropic’s Haiku and 1/8th of Gemini 3.1 Pro, though output pricing triples from the 2.5 version it replaces.

Why it matters: Cheap, fast models are becoming the real battleground in AI, and Flash-Lite’s benchmarks suggest Google isn’t sacrificing much intelligence to get there. But for all the benchmark strengths that the Gemini 3 has brought to the table, its consumer impact hasn’t felt on the same level as Anthropic and OpenAI in 2026.

QUICK HITS

🛠️ Trending AI Tools

• 👷 Viktor - OpenClaw for enterprise with 3,000+ integrations in Slack and Microsoft Teams, delivering board-ready PDFs, dashboards, and web apps*

• ⚡️ Gemini 3.1 Flash-Lite - Google’s fastest, cheapest Gemini 3 model for high-volume dev workloads

• 🧠 GPT-5.3 Instant - OAI’s ChatGPT default model update with fewer refusals and less hallucinations

• 📚 NotebookLM - New custom styles for infographics to turn info into visuals

*Sponsored Listing

📰 Everything else in AI today

Alibaba’s Qwen team faced a wave of departures, as staffers posted a coordinated “Qwen is nothing without its people” message echoing OpenAI’s 2023 mutiny.

Cursor CEO Michael Truell said its AI agent autonomously solved an open math research problem over four days, with stronger results than the official human solution.

Anthropic reportedly submitted a proposal for a $100M Pentagon drone swarm challenge before being barred from DoD work, as rival firms were selected instead.

xAI released a new ‘Beta 2’ version of Grok 4.20, with the update featuring better instruction following, reduced hallucinations, and more.

OpenAI VP of Research Max Schwarzer announced he is joining Anthropic, saying he is “looking forward (to) supporting my friends there at this important time.”

COMMUNITY

🤝 Community AI workflows

Every newsletter, we showcase how a reader is using AI to work smarter, save time, or make life easier.

Today’s workflow comes from reader Lenore B. in New Zealand:

"I had to prepare for a hearing where I would be cross-examined. I uploaded the documents (with confidential details removed) into an AI project and instructed it to be demanding and skeptical. I also asked it to review each session and identify what I was doing well. It stopped unproductive rumination, sharpened my reasoning, and gave me structured rehearsal. I walked into the hearing calm, clear, and confident."

How do you use AI? Tell us here.

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: Supreme Court ducks AI copyright question

• Read our last Tech newsletter: Apple’s budget iPhone goes AI-first

• Read our last Robotics newsletter: Bio-robotic spy roaches

• Today’s AI tool guide: Create killer thumbnails with Midjourney

• RSVP to our next workshop on March 12: AI coding bootcamp pt. 1

See you soon,

Rowan, Joey, Zach, Shubham, and Jennifer — the humans behind The Rundown

Good morning, tech enthusiasts. Apple’s entry-level iPhone just leveled up, without a price hike.

The iPhone 17e starts at $599 with flagship-grade A19 power, 256GB standard, and MagSafe back on the menu. And with Apple Intelligence built in, it’s the cheapest ticket into the Cupertino giant’s on-device AI push — resetting what “entry-level” even means.

In today’s tech rundown:

• Apple unveils iPhone 17e at $599

• MWC 2026’s wild new hardware

• Ultrahuman’s smart ring eyes U.S. return

• App that warns about nearby smart glasses

• Quick hits on other tech news

LATEST DEVELOPMENTS

APPLE

🍎 Apple unveils iPhone 17e at $599

Image source: Apple

The Rundown: Apple’s iPhone 17e is here, and $599 buys more than it used to. The new model, with preorders open on March 4, packs the A19 chip, doubles storage to 256GB, restores MagSafe, and ships with Apple Intelligence built in.

The details:

• The phone restores MagSafe and adds Qi2, boosting wireless charging to 15W, and comes in white, black, and a soft pink color (similar to that of iPhone 15).

• A new 48MP main camera delivers 2x optical‑quality zoom, improved low‑light performance, and Dolby Vision 4K/60 video capture.

• Apple bakes in its on‑device Apple Intelligence suite, enabling features like smarter photo cleanup, call screening, and translation.

• It is also rolling out a refreshed iPad Air alongside the 17e, a modest update built around Apple Intelligence with newer silicon and a longer support window.

Why it matters: Apple is turning its entry phone into the default Apple Intelligence device, pushing flagship‑class silicon, more storage, and modern charging down to $599. That raises the floor for what a midrange phone — and even Apple’s entry‑level iPad — has to deliver, and puts real pressure on Android makers.

MWC

🕹️ MWC 2026’s wild new hardware

Image source: Lenovo / Reve

The Rundown: At the Mobile World Congress in Barcelona, Lenovo’s foldable gaming handheld and Honor’s new Robot Phone showed how hardware makers are betting on new form factors to stand out in a saturated market.

The details:

• Lenovo unveiled the Legion Go Fold, a Windows-based gaming handheld whose foldable 7.7‑inch display can expand into an 11.6‑inch screen.

• Honor debuted its so-called Robot Phone, built around a motorized pop-out gimbal camera that tracks subjects and swivels like a tiny head.

• Xiaomi dropped the 17 Ultra, a Leica-co-engineered triple-camera flagship loaded with oversized sensors and on-device AI processing.

• The show’s wilder fringe included Tecno's magnetically modular ultra-thin concept and bendy phones designed to curl around your wrist.

Why it matters: The mobile industry has a complacency problem: upgrade cycles are slow, and spec bumps aren’t inspiring purchases. MWC’s answer is to make the hardware fun again, and the real story out of Barcelona isn’t any single device but a shared bet that AI‑native form factors are the best shot at boosting demand.

ULTRAHUMAN

👀 Ultrahuman’s smart ring eyes U.S. return

Image source: Ultrahuman

The Rundown: Ultrahuman is betting a redesigned, AI-powered smart ring can claw back the U.S. market — months after a bruising patent battle with Oura got its previous hardware blocked at the border.

The details:

• Priced at $479, the Ring Pro extends battery life to 15 days, pairs a redesigned sensor array with a dual-core processor, and can store 250 days of data.

• The ring ships with a Pro Charger packing 45 days of reserve power and supports Qi wireless charging — a first for the lineup.

• Ultrahuman is seeking U.S. Customs clearance after an ITC patent ruling halted imports, cutting off roughly 45% of its 700K daily active users.

• The company has also launched Jade, a real-time “biointelligence” engine that converts raw biometric data into actionable health guidance.

Why it matters: Smart rings are the sharpest testbed for AI-first wearables, and Ultrahuman is fighting to stay relevant as Oura and Samsung tighten their grip on the U.S market. With the Ring Pro still in regulatory limbo, the company is betting that new hardware and Jade’s real-time coaching will be enough to stand out.

SMART GLASSES

👓 App that warns about nearby smart glasses

Image source: Nearby Glasses

The Rundown: A new Android app called Nearby Glasses scans Bluetooth signals to flag camera-equipped smart glasses nearby — a direct counterpunch to the creeping normalization of what critics are calling “luxury surveillance.”

The details:

• Nearby Glasses scans ambient Bluetooth signals and alerts users when someone in range is wearing camera-equipped smart glasses.

• The app identifies manufacturer-specific Bluetooth signatures from devices, including Meta’s Ray-Ban smart glasses and Snap’s Spectacles.

• Creator Yves Jeanrenaud, a sociologist, built the app after researching how smart glasses were used in immigration raids and to harass sex workers.

• He told TechCrunch the app is an “act of resistance” against consentless ambient recording and is considering an iOS version.

Why it matters: Smart glasses don’t just record as phones do — they make recording invisible, collapsing the social cues that let people notice, object, or step away, and right now, there’s no legal framework requiring them to announce themselves. As surveillance tech gets sleeker, we’re likely to see more counter-tools like this app.

QUICK HITS

📰 Everything else in tech today

SoftBank-backed PayPay reportedly delayed its roughly $10B U.S. IPO amid weakening tech markets and rising geopolitical tensions.

Anthropic’s Claude surged to the No. 1 spot on Apple’s iPhone apps chart in the U.S., riding a wave of publicity after the Pentagon sidelined it from government AI work.

Paramount+ and HBO Max will be combined into a single streaming service once Paramount Skydance’s acquisition of Warner Bros. Discovery closes.

Instagram will soon notify parents who use its supervision tools when their teens repeatedly search for terms related to suicide or self-harm.

Elon Musk’s xAI is redeeming $3B in high-yield bonds at a premium as part of a plan to wipe out about $17.5B in debt tied to xAI and X, ahead of a blockbuster SpaceX IPO.

A U.S. federal court blocked Virginia’s law that would have limited under-16s to one hour a day on social media and required platforms to verify users’ ages.

GSK agreed to buy Canadian biotech 35Pharma for $950M in cash to bolster its pipeline of cardiopulmonary and obesity-related drugs.

MyFitnessPal snapped up Cal AI, the viral teen-built calorie-counting app with over 15M downloads and around $30M in annual revenue.

Chinese AI startup MiniMax more than doubled revenue to $79M in 2025, driven largely by overseas demand for its generative AI models and apps.

Amazon says three of its data centers in the Middle East were damaged by drone strikes tied to the escalating U.S.-Iran conflict, disrupting key AWS cloud services.

Xiaomi launched its new flagship 17 and 17 Ultra smartphones globally, keeping prices flat at $1,079 and $1,619 despite a surge in memory chip costs.

COMMUNITY

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: Supreme Court ducks AI copyright question

• Read our last Tech newsletter: Netflix exits $83B Warner Bros deal

• Read our last Robotics newsletter: Bio-robotic spy roaches

• Today’s AI tool guide: Transcribe any video for free with this local AI

• RSVP to our next workshop on March 12: AI coding bootcamp pt. 1

See you soon,

Rowan, Joey, Zach, Shubham, and Jennifer — The Rundown’s editorial team

Good morning, AI enthusiasts. Copyright law was written for a world where humans made things. AI broke that assumption… But the Supreme Court doesn't want to deal with it yet.

The court just passed on the biggest AI authorship case to date, keeping copyright law's "humans only" standard on the books. But with AI content now flooding every corner of creative industries, this fight is likely nowhere near finished.

In today’s AI rundown:

• Supreme Court ducks AI copyright question

• Anthropic wants your ChatGPT memories

• Transcribe videos for free with this local AI

• Alibaba's tiny AI tops models 13x its size

• 4 new AI tools, community workflows, and more

LATEST DEVELOPMENTS

AI & COPYRIGHT

⚖️ Supreme Court ducks AI copyright question

Image source: “A Recent Entrance to Paradise” by DABUS, Wikimedia Commons

The Rundown: The U.S. Supreme Court just passed on hearing the biggest case yet over whether AI art can be copyrighted, letting lower court rulings stand that say only humans can be authors — and kicking one of the defining IP questions of the AI era.

The details:

• The case centers on Stephen Thaler, a computer scientist who built an AI system called DABUS and sought copyright in 2018 for artwork it generated.

• The Copyright Office denied it, ruling only humans can be authors — a judge called it a "bedrock requirement" in 2023, and the DC Circuit agreed.

• Trump's DOJ also backed the Copyright Office, telling the court that copyright law was written for human creators and not machines.

• The appeals court noted Thaler could've claimed authorship himself rather than listing the AI, hinting the door isn't shut for AI-assisted works.

Why it matters: It’s wild to see an AI system making artwork years before the majority of the world. It also feels like an awkward ruling given the current state of AI content pouring into every creative sector — and one that will continue to be challenged by bigger entities like studios or creators that have serious money riding on the answer.

TOGETHER WITH YOU.COM

🧠 It’s not just about getting the prompt right.

The Rundown: When trying to spin up AI agents, companies often get stuck in the prompting weeds and end up with agents that fail to deliver dependable results. This ebook from You.com goes beyond the prompt, revealing the five stages for building a successful AI agent and why most organizations haven’t gotten there yet.

In this guide, you’ll learn:

• Why prompts alone aren’t enough and how context and metadata unlock reliable agent automation

• Four essential ways to calculate ROI, plus when and how to use each metric

• Real-world challenges at each stage of agent management and how to avoid them

If you're ready to go beyond the prompt, this is the ebook for you.

ANTHROPIC

🧠 Anthropic wants your ChatGPT memories

Image source: Anthropic

The Rundown: Anthropic launched a new tool that lets users port their saved preferences and context from other AI providers with a single copy-paste, coming during a surge in switches and new sign-ups as the company battles the Pentagon.

The details:

• Users copy a provided prompt into their current chatbot, paste the output into Claude's memory, and the switch kicks in within 24 hours.

• The tool pulls saved instructions, personal details, project context, and behavioral preferences from ChatGPT, Gemini, or Copilot in a single upload.

• Anthropic also opened Claude's memory feature to free users for the first time, letting everyone build persistent context across conversations.

• Claude Code also got a new auto-memory upgrade, now able to save project context, debugging patterns, and workflow habits on its own across sessions.

Why it matters: Memory upgrades are big news for getting the most out of any AI platform, but the timing isn’t subtle, given the current wave of consumer support for the company in the wake of the Pentagon’s ban. Giving all those new users an easy way to bring context over is a smart move for turning a viral moment into lasting retention.

AI TRAINING

📝 Transcribe any video for free with this local AI

The Rundown: In this guide, you will learn how to translate and transcribe any video file for free by running an AI model locally on your computer, without having to upload videos to sketchy, free transcription sites.

Step-by-step:

1. Open your terminal and run brew install ffmpeg then pip3 install -U openai-whisper. If not on Mac, you can find the commands you need here

2. ffmpeg is an open-source tool that lets you edit videos from your terminal, and openai-whisper is the OpenAI model that does the actual transcribing

3. To use it, just point it at any video file like this: python3 -m whisper your-video.mp4 --model base. It will run entirely on your machine for free

4. A 15-minute video should take ~2 minutes to transcribe, giving a .txt file and an .srt file with timestamps as the outputs

Pro tip: Whisper can also translate videos. You’d just have to add the language and translation flags to your command (more on it in the guide).

PRESENTED BY OPTIMIZELY

📈 See what real AI execution looks like

The Rundown: Most teams are stuck in AI pilot mode. Tomorrow, join Optimizely's free Agents in Action virtual event featuring Nathaniel Whittemore (host of The AI Daily Brief) and more — to see agentic AI working in live workflows.

You’ll learn how to:

• Operationalize AI across content, approvals & personalization

• Scale AI without breaking brand or compliance

• Put practical deployment frameworks to work across your org

Register before it’s too late.

ALIBABA

🧠 Alibaba's tiny AI tops models 13x its size

Image source: Alibaba

The Rundown: Alibaba released Qwen3.5 Small, a family of four new open-source AI models small enough to run on a laptop or phone — with the most powerful of the bunch outscoring an OpenAI model more than 13x its size on reasoning and knowledge.

The details:

• The Qwen3.5 Small Series spans four sizes, ranging from a 0.8B for phones up to 9B for laptops — all free for commercial use under an open-source license.

• The 9B outscored OpenAI's GPT-OSS-120B, which comes in at 13x its size on graduate-level reasoning and multilingual knowledge tests.

• All four models handle text, images, and video natively, with the 4B matching visual task scores that previously required models 20x larger.

• Elon Musk complimented the release, saying the models have “impressive intelligence density”.

Why it matters: These aren't replacing frontier models in capabilities, but for powering AI features inside mobile apps, reading documents offline, or handling quick visual tasks without a cloud bill, small models are where everyday adoption really takes off. Alibaba just made that layer even stronger and open for anyone with a laptop.

QUICK HITS

🛠️ Trending AI Tools

• 👷 Viktor— Free Openclaw secure, SOC 2-certified AI coworker for Slack and Teams that keeps your data private*

• 🤏 Qwen3.5 Small - Alibaba's tiny models that rival AI systems 13x their size

• 🍌 Nano Banana 2 - Google's new top-ranked AI image model

• 🧠 Claude - Anthropic’s AI assistant, with new memory features

*Sponsored Listing

📰 Everything else in AI today

AWS lost connectivity at a UAE data center after unidentified "objects" struck the facility amid the US-Iran conflict, with Anthropic’s Claude facing major outages.

OpenAI research scientist Aidan McLaughlin shared his views on the company’s Pentagon agreement, saying, “I personally don’t think this deal was worth it”.

The U.S. Treasury, Federal Housing Agency, and State Dept. became the first offices to move off of Anthropic, with Treasury Sec. Scott Bessent saying “no private company will ever dictate the terms of our national security.”

Apple announced the new iPhone 17e at $599, bringing Apple Intelligence to its most affordable iPhone with visual search, AI call screening, and live translation features.

MyFitnessPal acquired Cal AI, an AI calorie-counting app created by two 19-year-old founders that hit 15M downloads and $30M in annual revenue in under two years.

COMMUNITY

🤝 Community AI workflows

Every newsletter, we showcase how a reader is using AI to work smarter, save time, or make life easier.

Today’s workflow comes from reader Sam S. in New York, NY:

"I am a video editor, and I was tasked with creating a 2-minute trailer out of 6+ hours of interview footage of emotionally challenging subject material. I used Claude to transcribe the audio of all 6 interviews and asked it to pull the most impactful soundbites and create a time-coded 2-minute script.

After generating the script, Claude then drafted an Edit Decision List that I could import into Premiere and open as a timeline sequence, complete with Claude's edits and soundbites. This saved me hours worth of reviewing interview footage and helped with the emotional stress of watching the difficult material. The result was an impactful, dramatic trailer with a story arc as good as anything I could've scripted myself. "

How do you use AI? Tell us here.

🎓 Highlights: News, Guides & Events

• Read our last AI newsletter: OpenAI steps into Anthropic’s Pentagon void

• Read our last Tech newsletter: Netflix exits $83B Warner Bros deal

• Read our last Robotics newsletter: Bio-robotic spy roaches

• Today’s AI tool guide: Transcribe any video for free with this local AI

• RSVP to our next workshop on March 12: AI coding bootcamp pt. 1

See you soon,

Rowan, Joey, Zach, Shubham, and Jennifer — the humans behind The Rundown