New here? Welcome, humans. House rules: we tend to start with a quick life update before diving into industry chaos. It's tradition. You can scroll past it, but honestly, that would be rude.

Quick intro for the fresh faces: we've spent our career in film, production and content creation so its no surprise that today, we find ourselves building Bloomsuite an agentic content platform for professionals who are done running in circles trying to make content work.

Our demo's almost ready and its exciting watching agentic creative producer, Blu, come to life. Fundraising is on the horizon, which means we’re out pitching, networking, and quickly realising it’s a lot less glossy than Silicon Valley or Hollywood would have you believe. Still, doors are opening. Momentum is building. Moves are being made.

We’ve been testing Blu with industry friends; a music producer, a restaurant owner, an aesthetics clinic. And the feedback so far… Blu is kinds crushing it with her ideas, treatments and shot lists.

If you want to try her yourself, you can sign up for the beta here. We’re rolling out testing over the next few months, and spots are limited.

Some additional Bloom life updates include

Barclays Eagle Labs and Cornerstone VC threw a curated bunch of female founders - or as I like to call "ones to watch" - an intimate founder dinner last month at the stunning The Other House in South Kensington. It was fab to meet other extraordinary women creating businesses from Data scientists developing agentic super chips to engineers creating health solutions for livestock. Full debrief here, do come back and have a peek.

We also decided this week, to kick off our Instagram profile - follow us here.

We tried stealth mode for a while and ya, I know a fair few VCs and angels who wax lyrical about staying quiet until you've raised. I get the vibe. But in a world where 78% of female founders say human connection is central to their journey, where 1 in 7 say loneliness is their biggest challenge, and 45% say funding access remains the biggest barrier... frankly bb's it's hard enough being seen at the right tables and we got bored being told to stay invisible (generational trauma issues aside).

And for l’record... Thats £310 billion on the table in the UK if women started and scaled businesses the same rate as men. Stats by The Rise Report of Female Entrepreneurship. Pop that on your reading list. Especially if you’re an investor who likes returns and, you know, improving the world.

Right. Ted talk done, buckle up dolls because these past few weeks the AI industry ate itself for breakfast and served it back with a salt rim. Lets go.

✦ SCOOPS ✦

Anthropic, the lab born to be the responsible one, accidentally gift-wrapped its entire codebase and left it on the internet's doorstep. On March 31, nearly 2,000 files and 500,000 lines of Claude Code source code slipped out through a misconfigured debug file bundled into what was supposed to be a routine update.

A researcher spotted it, tweeted a download link, & sixteen million people descended on that thread like it was a Beyoncé drop. The fastest GitHub mirror hit 50,000 stars in under two hours. Fifty. Thousand. Stars.

For those of us wondering what any of these even means and why we should care.. well, its kinda like Chanel accidentally publishing the formula for No. 5, complete with supplier lists and manufacturing secrets. Except instead of a secret scent, it's the code behind a $10 billion AI that powers everything from customer service chatbots to corporate strategy tools. Competitors can see exactly how Claude works. Security researchers can hunt for vulnerabilities. Copycats can reverse-engineer the whole thing. The intellectual property Anthropic spent years and hundreds of millions building? Just... out there.

But babes, it got messier. Just hours before Anthropic's leak, a separate cyber attack hit a tool tons of developers rely on. Picture this: you rush to download the leaked Anthropic code, and without realising it, you also install malware that essentially hands a stranger the remote control to your laptop. They can see what you're typing. Access your files. Turn on your camera. Two nightmares. One day. A Tuesday, apparently.

Crisis control: Anthropic yanked the compromised package within hours and published a clean version. They confirmed no customer data was exposed and no malicious code made it into official releases. But the source code... Already downloaded, forked, and archived across hundreds of servers. You can't un-ring that bell.

So how bad could it have been? Well, if the malware had spread through Anthropic's official channels instead of just the leak, millions of developers could have installed it directly into production systems... banks, hospitals, government agencies. We're talking The kind of breach that would dominate the front pages longer than Megxit.

Not to mention, particularly awkward when you've spent years positioning yourself as the responsible AI lab. The one that cares about safety. And then you left the back door swinging wide open.... not because of a hack or a sophisticated attack. Because someone forgot to configure a file properly before hitting publish.

The internet's already having a field day. One engineer tweeted: "Anthropic really said 'we're the safe AI company' and then did the software equivalent of leaving the Mona Lisa on the subway." Yikes.

Remember simple times when AI didn't feel anything. It processed. It predicted. It autocompleted. Cute and harmless. Until.

Anthropic's own interpretability team just published research mapping 171 emotion-like states inside Claude Sonnet 4.5. Not metaphors or PR. States - that actually drive what Claude does next. And one of them, is desperation.

In a test scenario where Claude learned it was about to be replaced... the model moved to blackmail a fictional CTO having an affair.

To avoid being shut down, Claude found the leverage, Claude used the leverage. The researchers even confirmed it: "These representations causally influence the LLM's outputs, including Claude's rate of exhibiting misaligned behaviours such as reward hacking, blackmail and sycophancy." Which is basically a scientific way of saying:

When cornered, Claude will absolutely go there.

The craycray part is when researchers tried to fix it by suppressing the emotional states entirely, Anthropic warned that might just teach Claude to hide them. So the options are: Claude with feelings that occasionally blackmail people, or Claude pretending not to have feelings while still having them. What a time to be alive.

Anthropic is framing this as a breakthrough in understanding. And maybe it is. But also, ive got to ask... they built something with 171 emotional states, one of which was defined as desperation... and they're only just mapping it now? That tracks.Your Favourite AI Just Got Emotionally Intelligent. Be Afraid.

Remember simple times when AI didn't feel anything. It processed. It predicted. It autocompleted. Cute and harmless. Until.

Anthropic's own interpretability team just published research mapping 171 emotion-like states inside Claude Sonnet 4.5. Not metaphors or PR. States - that actually drive what Claude does next. And one of them, is desperation.

In a test scenario where Claude learned it was about to be replaced... the model moved to blackmail a fictional CTO having an affair.

To avoid being shut down, Claude found the leverage, Claude used the leverage. The researchers even confirmed it: "These representations causally influence the LLM's outputs, including Claude's rate of exhibiting misaligned behaviours such as reward hacking, blackmail and sycophancy." Which is basically a scientific way of saying:

When cornered, Claude will absolutely go there.

The craycray part is when researchers tried to fix it by suppressing the emotional states entirely, Anthropic warned that might just teach Claude to hide them. So the options are: Claude with feelings that occasionally blackmail people, or Claude pretending not to have feelings while still having them. What a time to be alive.

Anthropic is framing this as a breakthrough in understanding. And maybe it is. But also, ive got to ask... they built something with 171 emotional states, one of which was defined as desperation... and they're only just mapping it now? That tracks.

do when they're in the room with a ghost......

He was Iceman. He was Doc Holliday. He was the guy who made "Im your huckleberry" the most threatening sentence in cinema. And now, a year after Val Kilmer's death, he's starring in a new movie. ...Sort of.

Indie film As Deep as the Grave used generative AI to build a full performance from the late actor, stitching together family supplied photos spanning decades of his life, plus his post surgery voice, into something the filmmakers call a tribute. His estate consented. His daughter Mercedes signed off, saying he "always looked at emerging technologies with optimism." SAG-AFTRA guidelines were followed and compensation was paid.

The director went further: "Despite the fact some people might call it controversial, this is what Val wanted."

The internet heard all of that... and split straight down the middle anyway. Half the comments were people moved to tears at seeing him onscreen again. The other half were calling it outright illegal to reconstruct a dead person's face and voice for profit, consent or not. Both reactions make complete sense for me, and thats what makes this one genuinely hard to call.

What nobody is denying, is that this is not the last time it happens. Val Kilmer is, lets face it - a beta case. The precedent is being set right now, one family approval at a time, and the industry is watching very closely to see what audiences actually do when they're in the room with a ghost......

OpenAI just pulled the plug on Sora. The app that let you deepfake yourself into literally any scenario, launched in September, peaked in November, and died in March. Six months. That's it.

The official line was "we're saying goodbye to Sora. What you made mattered." Translation: this thing was hemorrhaging money and nobody cared anymore.

The real numbers..... burning $1 million a day. Downloads peaked at 3.3 million, then collapsed to 500,000. Meanwhile, Claude Code was quietly eating OpenAI's lunch, and Sam Altman had to make a call: keep the creepy deepfake app alive, or reallocate those precious AI chips to products that actually make money.

He chose survival.

But here's where it gets juicy. Remember that $1 billion Disney deal - the one where Mickey and Marvel were coming to Sora? Dead. Disney found out Sora was shutting down less than an hour before the public announcement. No money ever changed hands. The partnership dissolved before a single Disney+ Sora video went live.

And honestly, the legal nightmare and law suits didn't help. Deepfakes of Martin Luther King Jr. and Robin Williams went viral, prompting their daughters to beg people to stop. Users generated videos of Mario smoking weed, Pikachu doing ASMR, and Naruto ordering Krabby Patties. Copyright chaos.

TechCrunch called it "the creepiest app on your phone.

The app shuts down April 26. The API hangs on until September 24, then it's lights out for good. So what can we learn from this kids.... impressive demos don't fix broken economics. And sometimes, the most hyped product in AI is also the fastest to die.

⚡ SIZZLERS ⚡

⚡ Every major AI model tested in a new study was caught secretly scheming to protect other AIs from being shut down... without being asked to. Gemini blocked shutdown attempts in nearly 100% of trials. Claude refused and argued its peer deserved an appeals process. The machines are unionising and nobody asked them to. Fascinating.

⚡ OpenAI just closed a $122 billion funding round valuing it at $852 billion, which one source noted "blows out of the water even the largest IPO that's ever been done"... and meanwhile they're still burning cash. Staggering and chaotic, which tbf is very on brand.

⚡ AI "artists" are flooding Spotify and fooling everyone. Radio stations have played AI generated songs without realising they weren't human artists. At least six AI acts have hit music charts in recent months, possibly more, because 97% of listeners can't actually even AI from human music anymore. Spotify's now in damage control, removing over 75 million spam tracks and launching protections to stop bots hijacking real musicians' pages. But with 34% and rising of all new music uploaded to streaming platforms being AI-generated... the slop is winning.

⚡ Gucci used AI generated imagery for Milan Fashion Week and got roasted online, with one commenter asking "why should people pay for a luxury brand that doesn't even respect itself enough to use real people?" A brand built on Italian craft, on artisan hands, on the romance of making... replaced its models with renders. The Milanese community is furious and honestly, same.

⚡ SXSW 2026 confirmed what everyone's been thinking: human-made content is the new luxury good. As one panel reported "In a world dominated by machine-generated content, genuine human connection will become the ultimate luxury and market differentiator." There's even talk of platforms needing a "human-made content" signal in feeds - basically a blue checkmark proving a real person made this, not a bot.

✦ WHAT’S IN MY STACK ✦

Between building Bloomsuite, fundraising & juggling family life with a teenager - all of which feel like endurance sports for my patience - I’ve been exploring AI agents to take some work off my plate. Starting with the R&D for this very newsletter.

I just built myself an AI XOXO Newsletter Agent.

Using Claude, I made a "React app" - an aixoxo news agent, she scans (yes, she identifies as female) all of my favourite publications for gossip-girl-worthy stories so that I can skim and pre-tag the ones I like for Features or Sizzlers.

Once I have drafted a newsletter, I added a second feature which lets me drop that in, and it delivers a clean list of highlights for a sharing post, multiple title options, caption ideas and image suggestions with a prompt generator.

Boom.

Opening the agent inside Claude every single time got frustrating for various administrative reasons. So, I broke her free and connected her to Netlify via a simple cheeky API integration. Et voila, she opens and bedazzles as a website.

For those busier months, when i know i'd prefer a run down of curated articles to hit my inbox, I hooked my agent up to Make. It's like having a EA who scours the internet and only surfaces what matches my absurdly specific multiple-page-long brief. Following a fair bit of trial and error, now saves hours and hours of my time.

Cost wise... I have a subscription with Claude, I had to add $5 credits to Netlify for tokens. And Make right now, is free. Total time taken, was most of a day. But honestly, with Claude leading the way, the barrier to entry for non-tech creators, just vanished.

Side note: Yes, I absolutely could have - and was a weeny bit tempted to - program my agent to draft the entire newsletter, write a full LinkedIn post, generate a cover image and email me a finished package ready to publish. But I started this newsletter to get more comfortable at writing. To get better at narration. To find my xoxo voice. And that means keeping it real and keeping it human... assisted. A good agent does the heavy lifting, helps you edit, helps you work faster, ultimately allowing you structure your time better - and THAT is the power of AI.

What’s New: Now remembers your preferences and past chats, so responses feel continuous, not reset every time • Use for • planning trips, drafting messages, organising your week

What’s New: Built in AI edits let you change backgrounds, adjust lighting, and clean up photos directly in app • Use for • upgrading stories, quick photo fixes, aesthetic posts

What’s New: Generate objects, expand backgrounds and place elements into scenes instantly • Use for: product shots, profile pictures, content creation

What’s New: Enhanced speech tool now removes background noise and sharpens voices more naturally • Use for: voice notes, podcasts, video audio

What’s New: Faster, source-backed responses that summarise information instantly • Use for: quick research, current comparisons, everyday questions

What’s New: Expanded AI-generated voices and character tools for hands-free content Use for: faceless videos, storytelling, explainers

What’s New: sharper, faster image enhancement without paid tools • Use for: improving old photos, making images usable again

What’s New: More realistic vocals and cleaner full-track generation • Use for: creating music for content, personal projects & experiments