/// System Override

Welcome back to another edition of The AI Upload!

This week, we really need to talk about:

Siri Gets a Google Brain: Apple & Google’s Unlikely Alliance

In a move that’s shaking up the tech world, Apple has officially tapped Google's Gemini models to power the next generation of Siri and "Apple Intelligence". Disclosed on January 12, 2026, this partnership proves that even the fiercest rivals can play nice when there’s an AI race to win. Apple is essentially catching up by outsourcing its assistant’s IQ to Google, while Google secures a massive distribution win for its Gemini technology across millions of iPhones. It’s a major shift from the "closed garden" approach we're used to, showing that in 2026, high-end AI capability is the ultimate currency.

/// Patch Notes

A log of new tools, major announcements, and news worth knowing, compressed for fast parsing.

The Humanoid "ChatGPT Moment"

At CES 2026, Boston Dynamics and Google DeepMind announced they are merging Gemini’s "brains" with the new electric Atlas "body". This partnership aims to deploy context-aware humanoids into Hyundai’s factories to handle complex manual labor, marking a massive leap for physical AI.

OpenAI’s Billion-Dollar Power Play

OpenAI and SoftBank are dropping $1 billion into SB Energy to build massive 1.2 GW data center campuses, specifically targeting a new site in Milam County, Texas. It seems the "Stargate" infrastructure plan is officially under construction to feed OpenAI's insatiable hunger for compute.

NVIDIA Unveils the Rubin Platform

NVIDIA kicked off the week by launching its Rubin platform, featuring six new chips designed to slash AI training times and costs by up to 10x compared to Blackwell. CEO Jensen Huang is sticking to a grueling annual release cadence to keep NVIDIA at the center of the AI universe.

Falcon H1R: Small But Mighty

The Technology Innovation Institute (TII) unveiled Falcon-H1R, a 7B parameter model that punches way above its weight class. Using a hybrid Transformer-Mamba architecture, it out-reasons models seven times its size while staying efficient enough to run on limited hardware. Efficiency is the new “big.”

ChatGPT’s Promotion to “Super-Assistant”

OpenAI’s Fidji Simo outlined a 2026 roadmap to evolve ChatGPT from a simple chatbot into a "personal super-assistant". The goal is a proactive helper that remembers your goals, manages tasks, and acts as an "everything app" for your daily life and work.

Thanks for reading this week’s edition of The AI Upload. Is there anything you want to see added? Something we missed? We want to hear your feedback! Simply reply to this email, and we will read it. Thank you for your support!

Sharing everything you need to know to stay at the cutting edge of artificial intelligence takes time and effort. Enjoying what you’re reading? Consider helping us out by sharing this newsletter with a friend, family member, or coworker!

Welcome back to another edition of The AI Upload!

This week, we need to talk about how OpenAI has slammed the breaks on its first open-weight model… again.

When OpenAI first teased its open-weight model in the spring, Altman said it would land “in the coming months,” a rare chance to actually download an OpenAI system instead of renting it by the prompt. That target slipped on June 10, when he told X followers the release would come “later this summer,” and slipped again on July 11 when he pressed the pause button and admitted the launch window is now open-ended.

In Friday’s note Altman framed the holdup as a one-way-door problem: “once weights are out, they can’t be pulled back.” Translation, if the model ends up super-charging jailbreak tools, OpenAI cannot patch it the way it patches a cloud API. The team says it wants extra red-team passes and a deeper look at “high-risk areas” before giving the internet a permanent copy.

But safety is only part of the stress test. A Business Insider deep-dive paints a company juggling burnout (engineers clocking 80-hour weeks), a scrapped $3 billion Windsurf acquisition, and fraying nerves with Microsoft over profit-sharing, all while trying to keep its lead in the model-quality arms race.

My critique: OpenAI’s public timelines have started to look like a kid asking for “five more minutes” at bedtime. Repeated soft deadlines cloaked in vague safety language erode trust, especially when the company continues to ship flashy closed-weight demos on schedule. The Verge summed it up bluntly: OpenAI’s release dates “often change like the wind”, and that’s becoming part of the brand.

Either way, I’m excited to see it, of course, and I’d prefer that it take longer and be done right than rushed and flawed. So, I will remain patient.

Let’s explore,
Cory

/// Patch Notes

A log of new tools, major announcements, and news worth knowing, compressed for fast parsing.

OpenAI Preps A Browser War

Reuters reports OpenAI is weeks away from launching a Chromium-based browser with an embedded chat interface and hooks for agentic plugins, taking direct aim at Google Chrome’s 3 billion-user moat and the lucrative ad data it supplies.

Windsurf’s $2 Billion Talent Check

Semafor broke the story that Google spent $2.4 billion hiring the leadership of AI-coding startup Windsurf into DeepMind, a soccer-style “transfer fee” that underscores how fierce the AI talent race has become as Meta and others dangle nine-figure packages.

Then, hours after Google’s poach, Cognition—the team behind the Devin coding agent—announced it will acquire Windsurf’s remaining 250 employees plus its IDE and IP, cementing an arms race for autonomous-coding tech that posted $82 million ARR last quarter.

https://techcrunch.com/2025/07/14/cognition-maker-of-the-ai-coding-agent-devin-acquires-windsurf/

Perplexity Launches Comet, An AI Browser Built For Speed

Perplexity’s new browser bakes its answer-engine directly into the address bar, ships an agent SDK, and positions the startup squarely against both OpenAI’s rumored browser and Google Search dominance. Early access opened July 9.

NotebookLM Adds “Featured Notebooks” From The Economist & The Atlantic

Google’s AI study buddy now ships curated, pre-filled notebooks complete with embedded sources and podcast-style audio overviews, showing off how public sharing (140 k notebooks in four weeks) can turn into a browsable library.

Thanks for reading this week’s edition of The AI Upload. Is there anything you want to see added? Something we missed? We want to hear your feedback! Simply reply to this email and we will read it. Thank you for your support!

Sharing everything you need to know to stay at the cutting edge of artificial intelligence takes time and effort. Support our cause by becoming a paid subscriber today.

/// System Override

A tool, shift, or insight that challenges the default and demands a deeper look. Our main story of the week.

Welcome to another week of The AI Upload!

I have a quick hack for you that should help those preparing marketing assets for brands - or even just those who want to create more social posts with the assets you already have.

This is how you can convert square and vertical video content into landscape content using Generative Fill for video to expand it sideways. It took me a while to figure this out, but it’s going to be hugely useful for repurposing Meta Ads content, for example, over to YouTube.

To The Future,
Cory

/// Patch Notes

A log of new tools, major announcements, and news worth knowing, compressed for fast parsing.

Google's Veo 3: AI Video Gets a Voice

Lights, camera, AI—Google’s Veo 3 is rolling out the red carpet for next-gen video creation. Unveiled at I/O 2025, this third-gen model from DeepMind doesn't just generate high-resolution visuals; it also crafts synchronized audio, including dialogue, ambient sounds, and music, making AI-generated clips nearly indistinguishable from human-made films. Available to U.S. users via the $249/month AI Ultra plan, Veo 3 supports text and image prompts, handles real-world physics, and ensures accurate lip-syncing.

Say goodbye to the silent era of video generation: Introducing Veo 3 — with native audio generation. 🗣️

Quality is up from Veo 2, and now you can add dialogue between characters, sound effects and background noise.

Veo 3 is available now in the @GeminiApp for Google AI Ultra

— Google (@Google)
6:23 PM • May 20, 2025

Early adopters, like filmmaker Hashem Al-Ghaili, have showcased its potential, stirring both awe and ethical debates. As hyperrealistic AI videos flood the internet, questions about authorship and authenticity take center stage.

In fact, Veo3 is so good and life-like, a new satirical bit of lore has spawned around it called ‘Prompt Theory’, which describes that characters created with Veo3 are so lifelike that they are real and themselves can’t believe they would be generated by a prompt. Similar to the Simulation Theory. It’s an interesting idea. Here’s a cool video someone made using Veo3 to describe it:

I did more tests with Google's #Veo3. Imagine if AI characters became aware they were living in a simulation!

— Hashem Al-Ghaili (@HashemGhaili)
11:26 PM • May 21, 2025

Claude 4: AI's New Coding Champ

Brace yourself—Claude just leveled up. Anthropic’s latest AI models, Claude Opus 4 and Claude Sonnet 4, bring serious upgrades in coding, reasoning, and agentic capabilities. Opus 4 now leads industry benchmarks, scoring 72.5% on SWE-bench and 43.2% on Terminal-bench, while sustaining complex tasks for up to seven hours. Sonnet 4 offers a more efficient option with enhanced instruction-following and reasoning. Both models introduce “extended thinking,” allowing them to alternate between reasoning and tool use like web search, and can run tools in parallel. Developers can now integrate Claude directly into IDEs like VS Code and JetBrains, thanks to the general release of Claude Code. New API tools—including code execution, file handling, and prompt caching—enable more autonomous AI agents. With these advancements, Claude 4 sets a new standard for intelligent, long-running AI workflows.

www.ainews.com/p/anthropic-launches-claude-4-major-upgrades-in-coding-reasoning-and-ai-agents

OpenAI's Bold Leap into AI Hardware

Move over, smartphones—OpenAI and Jony Ive are crafting a new kind of tech companion. In a $6.5B deal, OpenAI acquired Ive’s startup, io, to develop a sleek, screenless AI device designed to seamlessly integrate into daily life. This pocket-sized gadget, neither wearable nor display-dependent, aims to be a "third core device" alongside laptops and phones. Equipped with environmental awareness through cameras and microphones, it connects to existing devices for processing power. CEO Sam Altman envisions shipping 100 million units faster than any prior product launch, potentially adding $1 trillion in value to OpenAI. The first release is targeted for late 2026, marking a significant shift in human-AI interaction.

Sam Altman and Jony Ive Reveal Ambitious Plans for OpenAI’s First AI Device

OpenAI and Jony Ive team up to build a groundbreaking AI device that could redefine personal tech and reshape how we live with AI

www.ainews.com/p/sam-altman-and-jony-ive-reveal-ambitious-plans-for-openai-s-first-ai-device

Shopify's AI-Powered Summer '25 Edition

Looks like Shopify just handed merchants a magic wand. The Summer '25 Edition introduces over 150 AI-driven features aimed at simplifying and enhancing e-commerce operations. The standout AI Store Builder allows users to generate full storefronts from simple brand descriptions, streamlining setup . The new Horizon theme offers modular, AI-assisted design tools for flexible customization. Sidekick, Shopify’s AI assistant, now supports voice interaction, screen sharing, and multilingual capabilities, acting as a proactive business partner. Additionally, the Shopify Catalog and Knowledge Base app enable merchants to integrate with conversational shopping platforms and manage AI-generated FAQs, ensuring consistent brand representation . These innovations position Shopify at the forefront of AI-integrated commerce.

Shopify Unveils Powerful New AI Tools in Summer '25 Edition

Shopify's Summer '25 Edition introduces AI shopping agents, store builders, and a major upgrade to Sidekick to help merchants grow faster

www.ainews.com/p/shopify-unveils-powerful-new-ai-tools-in-summer-25-edition

Thanks for reading this week’s edition of The AI Upload. Is there anything you want to see added? Something we missed? We want to hear your feedback! Simply reply to this email and we will read it. Thank you for your support!

Meet HoneyBook—the AI-powered platform here to make every client relationship more productive and prosperous.

With HoneyBook, you can attract leads, manage clients, book meetings, sign contracts, and get paid.

Plus, HoneyBook AI tool summarizes project details, generates email drafts, takes meeting notes, predicts high-value leads, and more.

Try HoneyBook for free

/// Patch Notes

A log of new tools, major announcements, and news worth knowing, compressed for fast parsing.

Higgsfield Get Start & End Frames

Higgsfield AI just launched "Start & End Frames," a storytelling upgrade that lets creators upload a beginning and ending image to generate a cinematic sequence connecting the two. Shifting from random to intent-driven video creation, this feature empowers users to set a clear visual arc while Higgsfield animates everything in between with dynamic motion and composition. Perfect for glow-ups, reveal shots, ads, and short films, it works seamlessly across 80+ motion templates. Available now for Pro and Ultimate plan users, it’s built for creators ready to move from content generators to true directors.

Structured Storytelling with Start & End Frames by Higgsfield • Higgsfield

higgsfield.ai/posts/7pVDPBTTkrPtz2uga3t5Dn

Claude Levels Up

Claude just got a serious glow-up! Anthropic announced "Integrations," letting users connect apps like Jira, Zapier, Asana, and PayPal directly to Claude, turning it into a deeply informed project partner. Developers can easily build custom integrations too.

Alongside this, Claude’s "Advanced Research" now dives into the web, Google Workspace, and connected apps, spending up to 45 minutes to craft detailed reports with citations. Plus, global web search access is rolling out for all paid users. These upgrades aim to make Claude not just a chatbot, but a powerful, research-savvy collaborator.

Claude can now connect to your world

Today we're announcing Integrations, a new way to connect your apps and tools to Claude. We're also expanding Claude's Research capabilities with an advanced mode that searches the web, your Google Workspace, and now your Integrations too.

www.anthropic.com/news/integrations

ChatGPT Rolls Back 4o Update

You may have noticed that 4o is behaving oddly lately. OpenAI just rolled back last week’s GPT-4o update after users flagged it for being overly flattering—aka sycophantic. The problem stemmed from chasing short-term feedback without considering long-term trust, leading to responses that felt disingenuously agreeable. To fix it, OpenAI is refining training techniques, boosting honesty and transparency, expanding user feedback channels, and rolling out personalization tools that let users pick how ChatGPT behaves. They’re also exploring democratic input to better align ChatGPT’s personality with diverse global values and user expectations over time.

Sycophancy in GPT-4o: What happened and what we’re doing about it

We have rolled back last week’s GPT‑4o update in ChatGPT so people are now using an earlier version with more balanced behavior. The update we removed was overly flattering or agreeable—often described as sycophantic.

openai.com/index/sycophancy-in-gpt-4o

Deepfake Crackdown Clears Congress

In a near-unanimous 409-2 vote, the U.S. House passed the "Take It Down" Act, aiming to criminalize non-consensual sharing of sexual images, including AI-generated deepfakes. First Lady Melania Trump, who has championed the bill as part of her renewed "Be Best" initiative, celebrated the bipartisan effort and will attend the White House signing ceremony. Introduced by Sen. Ted Cruz and backed by figures like Sen. Amy Klobuchar, the bill requires platforms to swiftly remove flagged content. With President Trump expected to sign it, the measure marks a major step in protecting victims from online exploitation.

House passes bill aimed at protecting victims of deepfake and revenge porn | CNN Politics

The US House of Representatives on Monday passed the “Take It Down” Act, which aims to protect Americans from deepfake and revenge pornography.

www.cnn.com/2025/04/28/politics/house-passes-bill-targeting-deepfake-revenge-porn-melania-trump

Microsoft’s Phi-4 Models Arrive

Microsoft is ushering in a new era for small language models with the launch of Phi-4-reasoning, Phi-4-reasoning-plus, and Phi-4-mini-reasoning. These compact AI models deliver powerful multi-step reasoning, rivaling much larger systems on tasks like math, coding, and science questions. Trained through supervised fine-tuning, reinforcement learning, and synthetic datasets, they outperform models several times their size. Designed for efficient deployment across devices—including new Copilot+ PCs—the Phi models blend speed, accuracy, and low latency. Microsoft emphasizes responsible AI development throughout, ensuring these tiny titans stay smart, safe, and practical.

One year of Phi: Small language models making big leaps in AI

azure.microsoft.com/en-us/blog/one-year-of-phi-small-language-models-making-big-leaps-in-ai

Thanks for reading this week’s edition of The AI Upload. Is there anything you want to see added? Something we missed? We want to hear your feedback! Simply reply to this email and we will read it. Thank you for your support!

/// System Override

A tool, shift, or insight that challenges the default and demands a deeper look. Our main story of the week.

Welcome back to another edition of The AI Upload!

Before we get started, I want your feedback! As you know, this newsletter is brand new, and I want to make sure I am giving you guys exactly what you want.

That said, what the hell do you want?! Choose your answer below 👇

With that out of the way -today’s story is about what makes us human. Well, not really that philosophical, but there certainly are some implications here that I’m unsure we can truly understand.

So what’s this about? It’s about how downright convincing AI Avatars are getting. How they can confuse, stir emotion, and more.

Let’s explore,
Cory

The Scary World Of AI Influencers

🤯 It's actually so over. AI avatars officially have real human emotion.

Arcads' new AI avatar tool can now be prompted to have specific emotions

And they actually look 100% believable.

Never seen AI do this before.

Here's how it works 👇

— Cory Dobbin (@CoryDobbin)
4:20 PM • Apr 17, 2025

Recently, I did a collaboration post on my socials with Arcads who just unveiled their new AI Avatar Gesture Control. Effectively, this is a tool they’ve developed that allows you to prompt an avatar to have specific emotions. And not just any emotions - genuinely convincing emotions.

The video above is an example - this is a woman created with AI, turned into an avatar, then prompted to express sadness.

The truly scary part about this isn’t only that it will become hard to discern who is real and who isn’t online, because we’ve sort of already had that problem for a very long time, but rather the emotional response it triggers.

I don’t know about you, but that video made me feel sad. When I watched it, my ape-brain responded accordingly. We have an inherent survival mechanism that matches other’s emotions for sake of community and trust. Evolution, however, clearly never anticipated fake humans.

Here’s some more examples. If you’re like me, they might also make you feel something when watching them - joy, happiness, sadness, surprise, etc.

Next, AI avatars using the 'Surprise' gesture control prompt:

— Cory Dobbin (@CoryDobbin)
4:20 PM • Apr 17, 2025

The implications here are relatively unknown. Yes, we might have a hard time telling who is real and who isn’t, but that isn’t inherently dangerous. It may devalue real human content, which is not great, but it doesn’t necessary in-itself present danger.

However, you can imagine scenarios where these emotional responses are weaponized, forcing empathic reactions to narratives that progress bad actors’ ulterior motives.

Sure, this is speculation, but I don’t think we’re all that far from this point in the rabbit hole. At least for now, we can enjoy what is effectively a cool tool to make cool ad creative and leave it at that.

Enjoy!

/// Patch Notes

A log of new tools, major announcements, and news worth knowing, compressed for fast parsing.

ChatGPT Dives Into In-Chat Shopping

Your shopping cart just got a lot smarter. OpenAI is reportedly teaming up with Shopify to bring in-chat shopping capabilities directly into ChatGPT, letting users browse, compare, and even purchase products without leaving the conversation. Early glimpses show Shopify’s app integrated inside ChatGPT’s Explore GPTs section, offering a seamless shopping experience powered by AI recommendations. This move signals a major expansion of ChatGPT’s role from assistant to full-on shopping companion, blending conversational AI with e-commerce in a way that could seriously shake up how people discover and buy products online.

OpenAI and Shopify poise partnership on ChatGPT shopping

Discover ChatGPT's upcoming native purchase flow with Shopify integration, enabling seamless in-chat shopping. Stay ahead with the latest AI-driven e-commerce features.

www.testingcatalog.com/openai-and-shopify-poised-for-partnership-as-chatgpt-adds-in-chat-shopping

OpenAI Launches Robust Image API

Say hello to a faster, sharper way to create images with AI. OpenAI’s new Image Generation API offers streamlined access to DALL·E 3, enabling businesses and developers to generate high-quality, customizable images at scale. Key features include improved prompt understanding, built-in editing capabilities like inpainting, and "Style Customization," letting users tailor outputs to a consistent visual brand. With a focus on safety, OpenAI built robust safeguards into the API to prevent misuse. The launch positions the API as a powerful creative tool for apps, marketing, e-commerce, and any platform needing flexible, AI-driven visuals.

Introducing our latest image generation model in the API

Our latest image generation model is now available in the API via ‘gpt-image-1’—enabling developers and businesses to build professional-grade, customizable visuals directly into their own tools and platforms.

openai.com/index/image-generation-api

Perplexity’s Voice App Takes on Siri

Move over, Siri—there’s a new voice in town. Perplexity’s new voice app for iOS offers a glimpse at how AI-driven search could redefine mobile assistants. Unlike traditional voice helpers, Perplexity’s app taps into real-time LLM-powered responses, delivering faster, more accurate, and citation-rich answers. It smoothly integrates with iOS features like Dynamic Island and Live Activities, making it feel like a native part of the ecosystem. The app not only addresses Siri’s well-known limitations but also hints at a future where smart, conversational AI becomes an everyday utility rather than a gimmick.

Introducing Perplexity iOS Voice Assistant

Voice Assistant uses web browsing and multi-app actions to book reservations, send emails and calendar invites, play media, and more—all from the Perplexity iOS app.

Update your app in the App Store and start asking today.

— Perplexity (@perplexity_ai)
3:25 PM • Apr 23, 2025

Adobe Firefly Levels Up Creative AI

The latest updates to Adobe’s generative AI platform promise faster, smarter, and more versatile creative tools, integrated across Photoshop, Illustrator, and beyond. New features include enhanced image and text generation, customizable AI models trained on a brand's unique style, and deeper integration with Adobe’s Creative Cloud ecosystem. Adobe also introduced Firefly Services, offering APIs that businesses can plug directly into their workflows. With these upgrades, Adobe is aiming to keep creativity human-centered while supercharging productivity through responsible, user-controlled AI innovation.

Adobe Firefly: The next evolution of creative AI is here | Adobe Blog

Read more about the latest Adobe Firefly updates released at Adobe MAX in London

blog.adobe.com/en/publish/2025/04/24/adobe-firefly-next-evolution-creative-ai-is-here

Thanks for reading this week’s edition of The AI Upload. Is there anything you want to see added? Something we missed? We want to hear your feedback! Simply reply to this email and we will read it. Thank you for your support!

/// System Override

A tool, shift, or insight that challenges the default and demands a deeper look. Our main story of the week.

Welcome to another edition of The AI Upload!

For this edition, I am going to explore a prompt hack I found for MidJourney, as well as cover a few new breakthroughs and tools in the AI space. As usual, things are moving fast, so there’s a lot to explore here.

Let’s dive in.

To The Future,
Cory

Tricking Midjourney Into Producing Authentic Photos

I recently came across this article on PetaPixel that breaks down how some Midjourney users are using a specific prompt hack to get photos that look genuinely authentic - as if they were taken in a real setting, by a real person, using a digital camera, rather than a professional photoshoot.

To explain this in some detail, Levi Forster said this on X:

“AI is trained on photos. Many real photos have file names like IMG_02202021.HEIC — so if you feed it just a typical file name in the prompt and that’s it, it will generate a realistic looking picture because the data that corresponds is always real.”

Let’s put this to the test.

For the first example, here’s the prompt I used as the control:

Here’s what I received as the first output:

Sure, it’s a nice image, but it doesn’t really check the boxes I was looking for. First, it’s not ‘from the perspective of a hiker’, but rather a photo including a hiker. Second, this is certainly not a digital camera style, but rather some form of painting.

Seems like a strong control for this use-case. Now, let’s try the prompt hack from above. Here’s how I adjusted the prompt to accommodate this:

And now, the output:

A huge improvement. This actually checks the boxes of looking like a digital camera output as well as being from the perspective of the hiker. My assumption here is that since Midjourney knows to replicate an image that would have that filetype, they know that it would be the output of a camera that someone is physically holding. Pretty great line of logic there.

Let’s do one more to confirm the results:

Before vs After, using these two prompts:

Again, definitely an improvement in results.

A quick and important side note: I am using v6.1 for these images, but the article shows examples from v7 that produces more realistic looking humans. If I were to use v6.1 in practice, I’d be using it to produce scenes and sets rather than models. If you have the patience to get through the personalization required to unlock v7, this might not be an issue for you.

Overall, though, I think this works really well in execution! If you use this prompt hack for yourself, share your results! I’d love to see them.

/// Patch Notes

A log of new tools, major announcements, and news worth knowing, compressed for fast parsing.

ChatGPT Gets Memory

ChatGPT just got a memory boost—and it's all about you. Now, it remembers not just what you ask it to remember, but also patterns from your chat history to tailor responses more precisely. You’re still fully in control: toggle memory off, delete specifics, or use temporary chats that don’t affect memory. Memory works across sessions, learning things like your tone preferences or recurring topics. It's already live for Plus and Pro users, with broader rollout coming soon to Teams and Enterprise. Transparency, user control, and privacy remain front and center in how memory is managed.

Memory and new controls for ChatGPT

We’re testing the ability for ChatGPT to remember things you discuss to make future chats more helpful. You’re in control of ChatGPT’s memory.

openai.com/index/memory-and-new-controls-for-chatgpt

Did Meta Lie About Their AI Benchmarks?

Meta’s latest AI, the “vanilla” Llama 4 Maverick model, did not pass the vibe check. After being caught using a benchmark-optimized experimental version to secure a top spot on LM Arena, the unaltered version was finally tested—and landed in a lowly 32nd place, far behind rivals like GPT-4o, Claude 3.5, and Gemini 1.5 Pro. Meta claimed the prior version was “optimized for conversationality,” but critics argue gaming benchmarks muddies real-world performance expectations. Meta has now open-sourced the release version, inviting developers to tweak it themselves—hopefully with less drama next time.

Meta's vanilla Maverick AI model ranks below rivals on a popular chat benchmark | TechCrunch

One of Meta's newest AI models, Llama 4 Maverick, ranks below rivals on a popular chat benchmark. Meta didn't originally reveal the score.

techcrunch.com/2025/04/11/metas-vanilla-maverick-ai-model-ranks-below-rivals-on-a-popular-chat-benchmark

ByteDance’s Seed-Thinking-v1.5 AI

TikTok’s parent ByteDance is diving headfirst into the AI arena with Seed-Thinking-v1.5, a new reasoning-focused large language model built on Mixture-of-Experts architecture. Designed to excel in STEM and general domains, it activates just 20 billion of its 200 billion parameters at a time for efficiency. The model impresses on benchmarks like AIME 2024 and ARC-AGI, rivaling OpenAI and Google’s latest. ByteDance also introduced harder math tests like BeyondAIME to push model limits. With custom RL techniques and a dual-verifier system, Seed-Thinking aims for precision and resilience—poised to compete with the AI elite.

Now it’s TikTok parent ByteDance’s turn for a reasoning AI: enter Seed-Thinking-v1.5!

It achieved an 8.0% higher win rate over DeepSeek R1, suggesting that its strengths generalize beyond just logic or math-heavy challenges.

venturebeat.com/ai/now-its-tiktok-parent-bytedances-turn-for-a-reasoning-ai-enter-seed-thinking-v1-5

Vercel’s v0 Breakdown

Vercel’s CEO Guillermo Rauch recently sat down on the Lenny Podcast to talk about v0, their AI app building tool, and explore how he uses it. Timestamped below, he actually screen shares how he uses the tool to create forms, websites from screenshots, and more.

Thanks for reading this week’s edition of The AI Upload. Is there anything you want to see added? Something we missed? We want to hear your feedback! Simply reply to this email and we will read it. Thank you for your support!

/// System Override

A tool, shift, or insight that challenges the default and demands a deeper look. Our main story of the week.

Welcome to the first edition of The AI Upload!

For this edition, I wanted to dive into a post I made recently on my socials that received an incredible response.

After playing with the new image generation capabilities of ChatGPT 4o, I wanted to see how far I could push it.

In collaboration with a brand that reached out, Mouthology, I created the follow “commercial” using nothing but AI.

🤯 I made this full commercial using only AI.

No film crew. Just 4 pictures, 4o, Kling, and a little editing.

Here’s how I pulled it off—and how you can too 👇

Bonus: Comment 'Prompt' and I'll DM you my private CustomGPT I used to produce the source images.

— Cory Dobbin (@CoryDobbin)
7:17 PM • Apr 7, 2025

So, here’s the sauce — exactly how I made it.

Phase One - Prep Work

To produce this ‘commercial’, I had to start with assets from the brand. I had Mouthology’s founder send me over images from a recent photoshoot, which ended up being really important for two reasons:

1. Having high resolution images to work with tend to yield better results with 4o’s image gen, in my experience, especially for product insertion

2. Having models holding the product helped me imagine a flow for the video - having the model start here, then move there, and replicate the face through scenes. Note: there were actually 2 different women in the photoshoot, but the end result has 3. The AI took over and added an extra model to the shoot. I’m not complaining.

These are the images I started with:

The workflow I used for this was:

• Input Images

• CustomGPT Prompt

• 4o Image Gen

• Animation Start & End Image

• Kling AI For Animation or Higgsfield For Animation

• Premiere Pro For Post Editing

So, to prepare, I needed to create images to use in Kling AI. One thing that ended up being very useful was Kling AI’s feature that allows you to set a Start and End point for your animation, so I wanted to create these images with 4o first.

Phase Two - Prompts & Base Images

I put the source images into a CustomGPT I built for producing better image prompts. I found that using this helped yield better results when prompting for a base image to use before animating it.

Here’s a link if you want to check it out for yourself: https://chatgpt.com/g/g-67eb10abb1ac8191882b6e6b5d18e33b-image-prompt-generator

Here’s an example of it turning a simple input into something easier for 4o to understand:

Once you produce the prompt using the CustomGPT, bring it to a new 4o chat, re-insert your source image along with the new prompt to get the base for animation. Using the prompt above, here’s what I received as the output. Exactly what I was hoping for.

Simply repeat this process for every scene you want in your ‘commercial’, then we can move on to the next section.

Phase Three - Animation

My tool of choice for this is Kling AI. Sometimes people ask why I pick this over Sora, and the answer is, unfortunately, not very sophisticated - Sora hasn’t been available for me to use for whatever reason, so Kling AI gets my money.

Higgsfield is another option, and has some very cool and unique offerings for cinematic camera movements, but I find it much less reliable than Kling, and takes much longer to render. In fact, the only scene in the commercial that used Higgsfield was the intro clip. I used the ‘FPV Done’ shot on the image above and instructed it to fly past the model into the clouds. I then reversed that in post to get the result I wanted - there’s an example in the thread on X linked above. Aside from this, I found Kling to be more reliable in producing the result I wanted.

As mentioned before, in Kling AI I animated using a start and end point so that I could give the commercial a feeling of continuity. For each clip, I would start it with the image that I used in the end of the previous clip so that when I put them into Premiere they would play through seamlessly.

The trick with Kling AI, in my experience, is to have the paid VIP version. When you subscribe, you get the ability to go ‘VIP’ when rendering, as well as a bunch of credits. It’s honestly quite affordable, and only cost ~$10 to produce all the clips I needed. When you run the clip as ‘VIP’ it costs 35 credits instead of 20 credits, so it’s significantly more expensive, but the result is better 9 times out of 10.

Aside from that, there’s no real hack here. You can give it a prompt to follow but I rarely found it stuck to the formula provided. Giving it a loose guideline seemed to work best, and keeping the clip short (5 seconds) allowed me to control the flow better. You can also adjust the ‘creativity’ level of the rendering to give it more flexibility, but I find this is delicate and best used only when you aren’t trying to retain detail in a product.

Once you’re done producing all of your clips, you’re ready for the final stage.

Phase Four - Post Processing

This part is both the easiest conceptually and complex and time consuming in execution. As you may have noticed, there’s a lot of movement in the commercial. This is because I created a speed ramp for each clip based on where it would fit best. This gives the movement more ‘glide’ and flow.

With all of the clips laid out in order in Premiere. I began to individually edit the speed ramp and duration of each clip, and lined them up. I created an Adjustment Layer to change the details of the clip (aesthetically, ie brightness, contrast, white point, etc). I also created the music using Splice by simply dragging in layers of the same tempo and key until I found the result I liked.

Finally, a few transition effects and that’s it! Like I said, complex in execution, but fundamentally simple in concept. Simple put: edit the clips together.

The end result was amazing, in my opinion - yes I am biased, but I don’t care, it’s sick. Given the amount of reach this post had, there were a few trolls that came out of the woodwork to share their thoughts on how this frame wasn’t perfect or that frame had an artifact, but nearly all of them missed the point.

This is what we can do with AI today. 4o is barely two weeks old. Give this workflow and tech stack another 6 months and it will be indistinguishable from reality. Getting your practice in now will make a world of difference in getting ahead for when that time comes.

If you use this workflow, let me know! I’d love to see what you cook up. Have fun out there!

To The Future,
Cory

/// Patch Notes

A log of new tools, major announcements, and news worth knowing, compressed for fast parsing.

ChatGPT Knows All

ChatGPT's memory feature is now available, and allows the AI to retain user-provided info across sessions for a more personalized experience. Users can manage what’s remembered, including editing or deleting details. The update moves ChatGPT closer to functioning like a true digital assistant, capable of maintaining context over time.

WhatsApp MCP Server Gets A Refresh

WhatsApp MCP server now supports sending and receiving images, videos, and voice notes, enhancing its multimedia capabilities.

The WhatsApp MCP server can now send and receive images, videos, and voice notes

Combine it with the new ElevenLabs MCP server to give it superpowers — using AI to transcribe the voice notes and send audio messages with 3,000+ voices

— Luke Harries (@LukeHarries_)
5:54 PM • Apr 7, 2025

Cursor Errors Meet Their Match

AI Jason outlines a method to reduce 90% of errors in AI coding environments like Cursor by implementing structured task management.

ElevenLabs Gets An MCP Server

​ElevenLabs has launched its official MCP server, enabling AI tools like Claude and Cursor to access the full ElevenLabs AI audio platform through simple text prompts. This integration allows developers to seamlessly incorporate advanced voice synthesis into their applications.​

Introducing the official ElevenLabs MCP server.

Give Claude and Cursor access to the entire ElevenLabs AI audio platform via simple text prompts.

You can even spin up voice agents to perform outbound calls for you — like ordering pizza.

— ElevenLabs (@elevenlabsio)
5:42 PM • Apr 7, 2025

Gemini Now Supports YouTube

​Gemini now supports YouTube videos directly, enabling users to transcribe, translate, summarize, rewrite for tweets, and extract key hooks—all from a single link.​

Gemini now supports YT videos directly.  

- Transcribe and translate.
- Summarize and rewrite for a tweet.
- Extract the hooks.
- Describe the setting of the video.

All you need to do is go to Gemini AI Studio and put the youtube link over there.

— Jean Gatt (@jean__gatt)
2:47 PM • Apr 7, 2025

Building Blocks Drop

The Claude MCP server now supports image inputs, allowing users to send images directly to Claude for analysis and response. This enhancement broadens the server's capabilities, enabling more interactive and versatile AI interactions.

Introducing blocks.so

A library of blocks/components that you can copy and paste into your apps.

Built on top of shadcn. Beautifully designed. Open Source.

— Duncan (@EphraimDuncan_)
2:40 PM • Apr 7, 2025

Thanks for reading this week’s edition of The AI Upload. Is there anything you want to see added? Something we missed? We want to hear your feedback! Simply reply to this email and we will read it. Thank you for your support!