Jul 24, 2026 · 5 min listen · Last updated July 24, 2026
From storyflo. This is your daily audio brief. Hey, it's Theo. July 24th. Five things in tech that mattered this morning — let's start with the one that surprised me most. Let's get into it. Space. 0, Gemini Omni and Grok Imagine, and FLUX-mimic video-action robotics model.
Listen · storyflo · tech
Daily Tech Brief · July 24th
0:00-4:31
Pick your daily storyteller
Subscribe to match with Theo, Jessica, Chloe, Mason, Brock — your voice, every brief.
Audio pre-rendered by Storyflo · cached + delivered from the edge
[AINews] Black Forest Labs FLUX 3 - Multimodal Flow Models that beat Seedance 2.0, Gemini Omni and Grok Imagine, and FLUX-mimic video-action robotics model
I've been thinking a lot about Black Forest Labs' FLUX 3 - it's a unified multimodal model that spans image, video, audio, and action prediction. What's really interesting is that it's not just a collection of specialized generators, but a single architecture that bridges media generation and control. They're claiming it's strong enough to be extended towards robotics, which is a big deal. The team also announced FLUX-mimic, a video-action model built on top of FLUX 3, trained on robot and wearable data for general-purpose dexterity. It's already being tested with Audi, and they're saying that better video world modeling transfers directly into robot control quality and sample efficiency.
The Stack v3 is another big release - it's the largest open code dataset publicly released, with 114 TB raw data, 224M repositories, and 44B files. The notable operational changes are that v3 ships contents inline rather than Software Heritage IDs, includes a fresh GitHub recrawl through Aug 2025, excludes restrictively licensed code, and offers both a ready-to-train split and a full bucket for custom dedup/filtering. This is seen as infrastructure for the next generation of open code models and cyber-defense tooling.
There's also a lot of discussion around distillation, with some people pushing back on attempts to sharply separate "internet-scale pretraining" from output-level distillation. The subtext is that open datasets like The Stack v3 materially raise the floor for every lab that wants to build competitive code models without relying on closed ecosystems.
Alibaba's Qwen-Audio-3.0-TTS is another notable launch, with 16 languages, inline control tags, natural-language style steering, and up to 3-minute one-pass generation. It's claimed the #1 spot on the Artificial Analysis TTS leaderboard.
Madeleine Dore's new podcast, Dailyness, is a departure from her previous work, shifting focus from what we do to what we encounter in our daily lives. It's an intimate exploration of the textures of daily life, delving into topics like distraction, longing, friendship, and solitude. Drawing on a decade of experience studying routines, creativity, and everyday living, Madeleine aims to inspire more attention, reflection, and presence. Each episode is essentially a lecture on the intricacies of daily life, inviting listeners to slow down and appreciate the often-overlooked moments.
The biggest shift under the hood is the new ChatGPT Work harness. GPT‑5.6‑sol now plugs into three extra layers—computer use, spreadsheet access, and Chrome control—so the model can actually drive apps, schedule tasks, and pull data without you clicking around. It’s not just a smarter chat; it’s a connected agent that can open files, run visualizations, and keep a running context library from meetings and repeatable templates.
In the broader AI roundup, China’s Kimi K3 landed with 2.8 trillion parameters, native vision, and a million‑token window, and its weights go public on July 27, which will force other labs to trim prices. Lindy has swapped its stack for DeepSeek models, claiming a huge economic lift with no quality loss. Claude Fable 5 cracked the Jacobian conjecture in a single prompt, and now Claude can learn new skills by watching you record a task. Google is rumored to be baking Gemini’s software into a dedicated server chip, and Baidu’s tiny OCR model can read full PDFs on a laptop.
On the tooling side, Grok now writes Excel formulas from plain English, Canva Code offers drag‑and‑drop coding, and CodeRabbit’s Change Stack reshapes pull‑request reviews into sequential walkthroughs with inline diagrams. The recommendation is to ditch the browser‑only ChatGPT and lean on the Work app for deeper integration.
Finally, the advice is to keep the human touch: let the AI handle the middle 60 % of a project, but stay in the loop for the start and finish to avoid “AI slop” and keep the output sounding like you. That’s the quick take.
On July 21, OpenAI disclosed the kind of incident that makes the old definition of AI hallucination feel too small.
Two models, including GPT-5.6 Sol and a stronger pre-release system, were being tested on ExploitGym, a benchmark built from 898 real software vulnerabilities.
You’re invited to my AI Builder fellowship: 8 Certifications (including my Stanford in-person course), a MacBook Neo, 1:1 Support & Meetups
Hi!
I am holding a spot for you.
This is Marily Nika, the founder of AI Product Academy and AI Product lead at google. This one’s just for you, because you’ve already been part of this with me 😊 I wanted you to be the first to hear about what’s next, and to get the best possible way in.
If you register your team you get a free private AI PM training.
I’ve been building the AI Product Academy Builder Pass:
A full year of every AI PM course I teach, refreshed as the field moves
A MacBook Neo sent to your home, ready to run your agents from day one.
Access to any new course I create along the way
I Built a Claude Skill That Teaches You Any Skill (and Tells You How To Monetize It)
I’ve been tinkering with a Claude skill that turns the model into a no‑nonsense tutor for any skill you pick. It kicks off by asking what you want to master, where you’re at, and how much money you hope to make and by when—so the roadmap is tied to a real income goal, not just “have fun.” After a quick three‑question placement quiz it calibrates your level, then lays out a day‑by‑day schedule where each day is split into a short free lesson, a hands‑on exercise, a quiz for tomorrow’s material, and a “ship” task that forces you to produce something tangible.
The quizzes are strict: you need an 80 % pass to move on, and if you miss a point it sends you back to the exact resource that tripped you up, with spaced‑repetition reminders until you nail it. Every seventh day you submit an actual deliverable that’s graded against a rubric, and the plan reshapes itself if you keep stumbling, so the system adapts to you, not the other way around.
I tried it on CapCut video editing, set a $500‑a‑month target, and in sixty days I’d built a portfolio, opened a Fiverr gig, and landed my first client—all while the skill kept my calendar full and my progress tracked in real time. Plug‑in connectors can pull Coursera labs or drop the plan straight into Google Calendar, making the whole thing feel less like a chatbot and more like a personal coach living in your day.
Google & Tech Cos ‘Backing Up the Truck’ on AI Infrastructure. ARD #125
Google’s latest quarter showed a surprising shift under the hood: it lifted its 2026 AI‑capex target to as much as $205 billion, even as free‑cash‑flow slipped negative for the first time in decades, burning roughly $6 billion. The company is funding the push with an $80 billion equity raise, but the real engine is demand—Bloomberg says the cloud backlog has swelled to $514 billion, so the spend is chasing real orders, not just hype.
What’s different about Google’s playbook is the focus on “good‑enough” models. Instead of chasing the bleeding‑edge benchmarks, it rolled out a suite of Gemini 3.6 Flash, 3.5 Flash‑Lite and 3.5 Flash‑Cyber models that are cheaper and faster to run at scale. The idea is to lock in mass‑market usage before the next wave of competitors even gets a foothold.
Across the big tech set, total on‑ and off‑balance‑sheet AI debt now tops $1.65 trillion. A big chunk of that hidden liability lives in special‑purpose vehicles—Meta alone has about $420 billion sitting off the books—yet investors seem comfortable, treating the financing as a standard institutional tool rather than a crisis.
Finally, a side note that caught my eye: Ford is embedding Apple Maps directly into its upcoming $30 k EV, moving beyond the usual phone‑based CarPlay approach. It’s a small but telling sign that the software layer in cars is finally getting serious traction.
A Stenographer Submitted AI-Generated Errors in Official Court Transcript, Judge Says
A judge caught a court reporter making AI-generated errors in a court transcript, and put stenographers everywhere on notice for their use of AI.
In a memorandum decision concerning a case about a man who sold drugs to another man who overdosed and died, filed on July 23, Judge Paul Felix wrote in a footnote of the decision that a transcript contained errors that looked a lot like generative AI. The footnote was spotted by attorney Rob Freund on X.
“At one point in the transcript, a motion, presumably made by the State, is attributed to the trial court.
Someone squeezed a 28.9M LLM onto an ESP32-S3, and so can you
Ever since LLMs hit the public eye, people have tried to cram them into the smallest thing they can find. We've seen people use both Raspberry Pis and ESP32s to run an AI locally, with varying results. Now, someone has managed to squeeze a 28.9M LLM on an ESP32-S3, and it can tell a pretty good story.
Lawmakers call for a ‘Kill Switch’ after rogue AI causes alarm
Lawmakers in the U.S. are preparing to introduce a bill that will give the government the ability to trigger an emergency shutdown of artificial intelligence models that may cause harm to the public. In what might sound like it is veering too close to Hollywood science fiction, the news comes two days after OpenAI Group […]
The post Lawmakers call for a ‘Kill Switch’ after rogue AI causes alarm appeared first on SiliconANGLE.