About This Week’s Covers

This week’s cover had to incorporate OpenAI’s new controversial image generation tool. There’s an uproar about copyright theft and AI ethics, so I figured the best artist to reference would be Banksy. I asked the image tool to recreate The Flower Thrower, but with the flowers replaced by a laptop to symbolize technological disruption.

Ironically, GPT refused to make the image. So, I gave it the original image and asked it to describe it for me. Then, I started a new chat, pasted the description, and asked it to change the details to a laptop computer and add my title. I fed that modified description to GPT-4o, which pretty much nailed it on the first try.

I used Claude 3.7 and Flux Pro Ultra to automate the creation of 40 category covers in one bulk command. Here are my favorite six from the batch. The goal is to create these quickly, not perfectly. Note that RAG stands for “retrieval augmented generation” (where AI references a document library to ground its answers). The fact that Claude conceptualized a person ripping pages to paste into the shape of the letters RAG is wild.

This Week By The Numbers

Total Organized Headlines: 505

This Week’s Executive Summaries

This week was a bit off-the-charts, and before I get into the summaries, I want to call out two terms that I think everyone needs to know: multimodality and interpretability. Keep these in mind as we talk through the summaries this week.

Multimodality is when a model can understand audio and images in full context. For example, if shown an image, a multimodal model will understand everything in the image, including the aesthetics, the objects, and the mood in a way that is hard to comprehend. This allows us to talk to or ask the model to modify parts of the image using plain language, and the model will understand what we mean. Since movies are simply a series of images, this means that a multimodal model can completely understand an entire film. Multimodality can include the understanding of an audio waveform. If given audio, a multimodal model will understand what is in that waveform without having to decode it. That means it could understand the tone and language being used in an audio file, but simultaneously parse background noise containing birds or a motor running or a song (think Shazam). Hearing and seeing in native modalities is a lot more powerful than I believe people realize. I think multimodality is one of the least celebrated and least feared capabilities of frontier models.

Interpretability is the computer-assisted human ability to understand and attribute the methods and elements referenced when a language model generates a response. Language models are trained on huge data sets, but they don’t memorize everything like a copy machine or hard drive. Rather, everything is written into abstract mathematical relationships, multidimensional values, and vectors.

When a frontier model is asked to respond to a request, it’s almost impossible to back out the training materials that would’ve contributed to the response. Sort of like a human mind. We move through life accumulating experiences and inputs. If I were to have an emotional response to something that someone says because of an assumption that I was superimposing on that person’s intent, I’m not sure I could go back and tell you the exact moment in my life that caused the emotional reaction. Or if you asked me to draw a picture of a pilot flying a plane, I’m not sure I’d be able to assign value to the vast influences over time that contributed to the style or composition of the way I would draw a plane.

On one hand, the ethics of an AI model training on copyrighted material seem very straightforward, but on the other hand, it’s hard to attribute at any granular level what a model is doing under the hood when it responds to any given request. The study of this is called interpretability. I feel that until we solve interpretability, we won’t be able to effectively attribute any training material influence on an AI model’s output. Marketers have struggled with multitouch campaign sales attribution for decades. I’m not sure how we will solve this one when it comes to training attribution for any given AI output.

This week I organized 506 headlines. 21 merited executive summaries.

The viral story of the week was OpenAI integrating image generation directly into GPT-4o. This is a prime example of a multimodal feature. It resulted in a surge of people using the tool to generate all sorts of stylized memes, but it also created a backlash and debate about the ethics of training. One particular type of anime style by a studio named Studio Ghibli ran rampant. Since the feature is multimodal, you can give it a picture of anything and ask it to re-style it in any fashion. You can ask it to combine images. You can ask it to add or remove objects. You can tell it to add text to it. It’s almost like having Adobe Photoshop with voice mode. I’ll provide a bunch of examples below. Here’s a post showcasing testing with personal imagery.

A few months ago, Anthropic released a structured method of allowing AI models to talk to third-party systems. This is called the “Model Context Protocol” or MCP for short. This week OpenAI announced that it would adopt Anthropic’s standard. This is a big deal for people who want to create agents that can execute tasks across a variety of third-party systems. It’s one step closer to being able to ask your phone to order something from DoorDash via conversational interface, never touching your phone, all the way through order completion.

Google came out of nowhere and launched a new version of their Gemini model, and has completely destroyed the competition. In any other given week this would’ve been the top story by far. Gemini is now at the top of the leaderboards in almost every category, with 40-point leads in some cases. I think next week this will get more attention. My sad opinion is that Google kills every product they touch, but AI has long been their superpower under the hood.

An image generation tool called Ideogram released a new version, which is very powerful and a major update. It’s being overshadowed by OpenAI, but in any other given week, it would be a big story. Image generation is getting cheaper and better. In the case of Ideogram, which could already make complex images with a lot of text, they have now boosted their photorealism to compete more with Flux and MidJourney.

A school in Austin, Texas claims to have reached the top 2% in academic test scores by using an AI tutor.

The State of Pennsylvania completed a government AI pilot program where 178 employees averaged a daily savings of 95 minutes by using ChatGPT.

Researchers from Harvard conducted a randomized trial with 776 Procter & Gamble employees and found that a single person can work as well as a two-person team if given AI tools. They also found that specialists produced work that was more reflective of a holistic team approach than when a specialist worked alone.

Retailer H&M is creating digital twins of 30 of its fashion models. The technical term for this is a LoRA, but what it means for marketing is that these models have licensed their likeness and the brand team at H&M can simply prompt these clones into any context or condition and use them for campaigns in the future without having to do a photo shoot. The examples are compelling and essentially indistinguishable from photos.

A trend in the last two weeks is developers releasing text files with detailed language model instructions in text. Last week, Stripe did this with its payments tool. An instruction file is similar to an API, but it is tailored towards language models. By giving a model the text file, a developer can load everything the language model needs to know about a system before building for it.

DeepSeek released a new update this week. Practically speaking, China has closed the AI development gap with the US to just three months in key areas.

Perplexity is proposing bringing TikTok into an open-sourced algorithm hosted in the United States.

Lockheed Martin and Google announced an integration of Google’s AI tools into Lockheed’s factory systems.

A relatively unknown image generation tool called Reve has come out with a model that is orders of magnitude cheaper than the top image tools with similar quality. It’s a bit of a ‘DeepSeek moment’ for image generation, although it’s not open-sourced. I predict we will see more about Reve next week once the OpenAI image buzz settles down.

OpenAI is close to completing a $40 billion funding round led by SoftBank that would nearly double OpenAI’s valuation in just six months.

Chinese AI agent model, Manus, is seeking tens of millions in funding from US venture capitalists. Similar to Perplexity, Manus doesn’t create AI models, but rather leverages them to build products. This is an interesting trend to keep an eye on.

Researchers have developed an AI model that can identify endometrial cancer in tissue samples with a 99.2% accuracy rate. This is significantly better than the current best-in-class methods which are at 80%.

There are a few more executive summaries in addition to these and also 5 separate compelling visuals of the week to check out. For the real gluttons for punishment, I also selected the remaining top 38 headlines of the week at the bottom. As usual, there’s no sign of slowing down.

GPT-4o Brings Text-Accurate Image Generation to OpenAI’s Flagship Model
OpenAI has integrated image generation directly into GPT-4o, creating a visual tool that crushes text rendering, handles complex instructions, and can understand and leverage contextual knowledge. The system maintains consistency across multiple image refinements and handles up to 20 distinct objects in a single scene. Unlike previous image models focused on artistic imagery, 4o prioritizes functional visuals for communication from diagrams and infographics to interfaces with accurate text. The multimodal model also learns from uploaded reference images, applying their characteristics to new generations while maintaining coherence through conversation-based refinements (and that’s what’s taken the internet by storm this week).

“4o-image is the first time many people will start to think of the post-reality-filter stage rolling out over the next few years. one’s reality is whatever they want it to be (ghibli or pokemon or lotr or…), and as each human finds that which they truly desire, they then” / X https://x.com/nearcyan/status/1905219687547621740

“Even with multimodal image generation, ChatGPT only gets 1 out of 3 of the hard challenges of image creation: ✅Horse riding an astronaut ❌Clock reading 5:30 ❌Full to the brim glass of wine Presumably, the training image data is mostly not-full glasses and clocks set to 10:10 https://x.com/emollick/status/1904710558919651819

Introducing 4o Image Generation | OpenAI https://openai.com/index/introducing-4o-image-generation/

“New image model from OpenAI is pretty good at UI stuff. https://x.com/skirano/status/1904609866099933272

“// i lead model behavior at openai, and wanted to share some thoughts & nuance that went into setting policy for 4o image generation. features capital letters (!) bc i published it as a blog post: — This week, we launched native image generation in ChatGPT through 4o. It was” / X https://x.com/joannejang/status/1905341734563053979

“This is a pretty cool feature that got drowned by Ghiblification hype: you can ask 4o image gen for transparent backgrounds! Should be super useful for creating all kinds of assets.” / X https://x.com/giffmana/status/1905407013103747422

“The instruction following capability of gpt-4o native image gen is pretty incredible. Nothing comes close to this, how could every other image gen model miss the mark” / X https://x.com/abacaj/status/1905075484892836308

“GPT-4o image gen is fun. It can do photorealistic & stylized renditions while keeping consistent structure and pose. “Make an image of a UFO parking not allowed sign with a UFO conspicuously parked in that spot, with a grey alien arguing with a police officer in front of it” https://x.com/bilawalsidhu/status/1904627124109271238

“we are launching a new thing today—images in chatgpt! two things to say about it: 1. it’s an incredible technology/product. i remember seeing some of the first images come out of this model and having a hard time they were really made by AI. we think people will love it, and we” / X https://x.com/sama/status/1904598788687487422

“💥 Today we’re rolling out a *major* update to image generation in ChatGPT! The model is now quite good at following complex instructions, including detailed visual layouts. It’s very good at generating text. It can do photorealism or any number of other styles. Btw—if you” / X https://x.com/kevinweil/status/1904595752380465645

“Entire ComfyUI workflows just became a text prompt. Open an image in GPT-4o and type “turn us into Roblox / GTA-3 /Minecraft / Studio Ghibli characters” https://x.com/bilawalsidhu/status/1904908540063424898

“My “otter on a plane using wifi” benchmark has now been saturated by ChatGPT 4’s new image generator: “an otter on an airplane using wifi, on their laptop screen is image generation software creating an image of an otter on a plane using wifi,” first try https://x.com/emollick/status/1904943271282934215

“”ChatGPT show me a photorealistic drone shot of a fantasy city, the walls are white alabaster streaked with gold, while massive brass trellises built into the towers allow vines of maroon to climb them. At the center is a tower with a balcony and a hooded figure on it” “in the https://x.com/emollick/status/1904720324530254049

OpenAI Adopts Anthropic’s Model Context Protocol
Big news for connecting AI models to third party systems (aka conversational DoorDash ordering). OpenAI will support Anthropic’s Model Context Protocol (MCP) across its products, including the ChatGPT desktop app. Sam Altman announced the move on X, citing user enthusiasm for the open-source standard. MCP creates connections between AI models and data sources, allowing AI assistants to access information from business tools, databases, and content libraries. The protocol, which Anthropic open-sourced, enables developers to build two-way connections through “MCP servers” and “MCP clients.” Several companies including Block, Apollo, Replit, Codeium, and Sourcegraph have already integrated MCP support into their platforms. Anthropic’s chief product officer Mike Krieger welcomed OpenAI’s adoption, noting that MCP has become a “thriving open standard with thousands of integrations.”

“people love MCP and we are excited to add support across our products. available today in the agents SDK and support for chatgpt desktop app + responses api coming soon!” / X https://x.com/sama/status/1904957253456941061

OpenAI adopts rival Anthropic’s standard for connecting AI models to data | TechCrunch https://techcrunch.com/2025/03/26/openai-adopts-rival-anthropics-standard-for-connecting-ai-models-to-data/

“Great to see OpenAI adding support for MCP! It’s amazing to think that in less than 4 months, MCP has gone from just an idea we had at Anthropic on how to make integrations easier for devs to the industry standard for all AI app integrations.” / X https://x.com/alexalbert__/status/1904965223448006805

Google Launches Gemini 2.5 Pro, Blows Everyone Out of the Water and Claims Top Spot in AI Benchmarks
Any other week, this would be the top story. Google has launched Gemini 2.5 Pro, a “thinking model” designed to tackle complex problems through improved reasoning capabilities. The model currently leads several major AI benchmarks by significant margins, with a 40-point ELO advantage (!) on the LMArena leaderboard. Gemini 2.5 Pro shows particular strength in coding, STEM subjects, multimodal reasoning, and creative writing tasks. The model is available immediately through Google AI Studio, with API access for developers. Google emphasizes the model’s ability to “think through” problems before answering, potentially offering more thorough solutions compared to competing AI systems.

“A mere ~16 point jump on https://x.com/OriolVinyalsML/status/1904920302053650713

“🚨 Gemini 2.5 Pro Exp dropped and it’s now #1 across SEAL leaderboards: 🥇 Humanity’s Last Exam 🥇 VISTA (multimodal) 🥇 (tie) Tool Use 🥇 (tie) MultiChallenge (multi-turn) 🥉 (tie) Enigma (puzzles) Congrats to @demishassabis @sundarpichai & team! 🔗 https://x.com/alexandr_wang/status/1904590438469951873

“Gemini 2.5 Pro Experimental on Livebench 🤯🥇 https://x.com/OfficialLoganK/status/1904925675892728179

“Introducing Gemini 2.5 Pro Experimental! 🎉 Our newest Gemini model has stellar performance across math and science benchmarks. It’s an incredible model for coding and complex reasoning, and it’s #1 on the @lmarena_ai leaderboard by a drastic 40 ELO margin. Only a handful of https://x.com/OriolVinyalsML/status/1904583691566727361

“Gemini 2.5 Pro is an awesome state-of-the-art model, no.1 on LMArena by a whopping +39 ELO points, with significant improvements across the board in multimodal reasoning, coding & STEM. You can try it out now in AI Studio https://x.com/demishassabis/status/1904587103805006218

“Today we are launching 2.5 Pro! I think it’s the best model in the world. State-of-the-art reasoning and great vibes (+39 ELO gap on lmsys!) 2.5 Pro improves in coding, stem, multimodal, instruction following, and lots more. Available in AI Studio & the Gemini App! https://x.com/jack_w_rae/status/1904583894458110218

“Introducing Gemini 2.5 Pro, the world’s most powerful model, with unified reasoning capabilities + all the things you love about Gemini (long context, tools, etc) Available as experimental and for free right now in Google AI Studio + API, with pricing coming very soon! https://x.com/OfficialLoganK/status/1904580368432586975

Gemini 2.5: Our newest Gemini model with thinking https://blog.google/technology/google-deepmind/gemini-model-thinking-updates-march-2025/

“Gemini 2.5 now does a much better job than most AIs writing plots & characters, thanks in part to its ability to “think” through details in advance. A fun test: “play a game of Fiasco (think a Coen Brothers movie, but as an RPG), make it interesting in terms of plot and prose” https://x.com/emollick/status/1904656593083396541

“Deep Dive Video: Complex image editing used to take hours — now Google’s Gemini 2.0 turns advanced ComfyUI & Photoshop workflows into simple text prompts. Here’s exactly how to try it (completely free). Chapters: 00:00 Conversational Editing with Google’s Multimodal AI 00:53 https://x.com/bilawalsidhu/status/1903172149764034605

“Google’s new Gemini 2.5 Pro Experimental takes the #1 position across a range of our evaluations that we have run independently Gemini 2.5 Pro is a reasoning model, it ‘thinks’ before answering questions. Google has released it as an experimental API in AI Studio only, and has https://x.com/ArtificialAnlys/status/1904923020604641471

“BREAKING: Gemini 2.5 Pro is now #1 on the Arena leaderboard – the largest score jump ever (+40 pts vs Grok-3/GPT-4.5)! 🏆 Tested under codename “nebula”🌌, Gemini 2.5 Pro ranked #1🥇 across ALL categories and UNIQUELY #1 in Math, Creative Writing, Instruction Following, Longer https://x.com/lmarena_ai/status/1904581128746656099

Google is rolling out Gemini’s real-time AI video features | The Verge https://www.theverge.com/news/634480/google-gemini-live-video-screen-sharing-astra-features-rolling-out

Ideogram Launches Version 3.0 of Its AI Image Generator With Enhanced Realism and Design Tools
Another example of any other week this would be huge. Even through OpenAI’s image tool went viral, Ideogram made a huge leap in quality. Ideogram has released version 3.0 of its text-to-image platform, featuring improved photorealism, text rendering, and style consistency. The updated model outperforms competitors in human evaluations across various capabilities. Users can now upload up to three reference images to specify styles or explore billions of preset options via a new “Random style” feature. The platform aims to make professional-quality design accessible to small businesses by enabling quick generation of logos, promotional materials, and product photography. Ideogram 3.0 is available on their website and iOS app.

Ideogram 3.0 https://about.ideogram.ai/3.0

Texas School’s AI Tutor Propels Student Scores to National Top 2%
Alpha School in Austin integrates AI tutors for personalized two-hour daily academic sessions, allowing students to spend remaining school hours developing public speaking, financial literacy, and teamwork skills. Test results place students in the country’s top 2% academically. The school claims their method enables students to complete core academics in just three hours daily, freeing time for passion projects. Ethan Mollick tweeted that he thinks they are potentially exaggerating the impact of AI and is doubtful of the results.

“As someone who has spent a lot of time thinking and building in AI education, and sees huge potential, I have been shown this headline a lot I am sure Alpha School is doing interesting things, but there is no deployed AI tutor yet that drives up test scores like this implies. https://x.com/emollick/status/1904193352045596745

AI ‘tutor’ boosts Texas private school test scores to top 2% nationally | Fox News https://www.foxnews.com/media/texas-private-schools-use-ai-tutor-rockets-student-test-scores-top-2-country

Anthropic Researchers Work to Reveal How Claude Actually “Thinks”
Anthropic scientists have developed a new “AI microscope” that traces Claude’s internal thought processes, uncovering surprising insights about how language models work. The research shows Claude thinks in a universal language across different human languages, plans rhyming poetry words in advance, and sometimes provides plausible-sounding but fabricated reasoning when challenged. This transparency effort demonstrates Claude doesn’t simply predict single words but engages in sophisticated planning and conceptual processing, though researchers note their methods capture only a fraction of the model’s total computation. This work aims to ensure AI systems remain reliable and aligned with human values as they grow more capable.

Tracing the thoughts of a large language model \ Anthropic https://www.anthropic.com/research/tracing-thoughts-language-model

“Anthropic continues to do some of the best AI interpretability research out there” / X https://x.com/iScienceLuvr/status/1905354818451112352

Pennsylvania Adopts ChatGPT for State Employees, Reports Saving 8 Hours Per Week Per Employee
Pennsylvania has completed a government AI pilot program where 175 state employees used ChatGPT in their daily work. According to Governor Josh Shapiro, the tool saved workers an average of 95 minutes per day—equivalent to 8 hours weekly or 30 hours monthly. Shapiro emphasized that Pennsylvania is focusing on “responsible, ethical use” of generative AI, with the primary goal of freeing up employee time to better serve residents.

“ChatGPT for Pennsylvania. Saves government employees 8 hours a week each:” / X https://x.com/gdb/status/1903278608673804573

“175 state government employees had the chance to participate in our first-in-the-nation pilot with @OpenAI to use gen AI to better serve Pennsylvanians. ChatGPT saved employees nearly 95 minutes a day — nearly 8 hours a week or 30 hours a month — giving them more time to get https://x.com/GovernorShapiro/status/1903152149879529589

AI Teammates Match Human Collaboration in P&G Study, Boost Performance and Satisfaction
Researchers from Harvard’s Digital Data Design Institute conducted a randomized trial with 776 P&G professionals, finding individuals using AI performed as well as two-person teams without AI. Teams with AI access produced the highest-quality solutions overall and were significantly more likely to create exceptional work. AI users completed tasks 12-16% faster while creating more detailed solutions. Surprisingly, AI eliminated traditional expertise boundaries. For example, commercial specialists and R&D experts produced similarly balanced solutions when using AI, regardless of their background. Workers reported more positive emotions and less anxiety when collaborating with AI tools. The researchers suggest organizations should view AI not just as a productivity tool but as a teammate that can democratize expertise and transform work structures.

“🚨in our new paper we ran an experiment at Procter and Gamble with 776 experienced professionals solving real business problems. We found that individuals randomly assiged to use AI did as well as teams of two people and were happier as well. Lots more in the link in the reply! https://x.com/emollick/status/1903441560525529527

“In our new paper on using AI solving real business problems at Procter and Gamble where we found AI helped individuals, even inexperienced ones, perform as well as teams of two, we gave participants a set of default prompts to work with. Here some are: https://x.com/emollick/status/1904340100412313631

PwC Launches Enterprise AI Operating System
PwC introduced agent OS, a unified orchestration framework for enterprise AI that connects and manages intelligent agents across business systems. The platform integrates agents from various sources into cohesive workflows, regardless of which platform they’re built on. Features include drag-and-drop interface, cloud-agnostic deployment, and compatibility with major tech ecosystems including Anthropic, AWS, Microsoft, Google Cloud, and others. The system addresses interoperability challenges while maintaining governance controls. Early adopters report significant improvements: a technology company reduced call center phone time by 25%, a hospitality company cut review times by 94%, and a healthcare organization improved access to clinical insights by 50% while reducing administrative work by 30%.

PwC launches AI Agent Operating System for enterprises: PwC https://www.pwc.com/us/en/about-us/newsroom/press-releases/pwc-launches-ai-agent-operating-system-enterprises.html

H&M Launches AI Digital Twin Models with Consent from Real Models
H&M is creating digital twins of 30 real fashion models for upcoming ad campaigns. The technology works by feeding multiple photos of each model into AI tools that generate realistic images that can be posed and styled digitally. This approach eliminates the need for physical photo shoots with photographers, makeup artists, and other production staff. H&M emphasizes that models will retain ownership rights to their digital twins and receive compensation when these AI versions appear in ads. While this arrangement protects the models involved, questions remain about the broader impact on fashion industry professionals whose roles might be diminished. The company acknowledges they’re still developing their approach to compensation and usage rights in this evolving space.

Clothing Giant H&M Will Use Models’ AI-Made Digital Twins, Consent Included https://www.inc.com/kit-eaton/clothing-giant-hm-will-use-models-ai-made-digital-twins-consent-included/91166352

Plain Text API Docs Help AI Models Integrate More Easily
Developers are now offering API documentation specifically formatted for AI consumption with new “/llms.txt” and “/llms-full.txt” files. These plain text resources allow large language models to better understand and integrate with APIs without requiring human interpretation. Professor Ethan Mollick notes this approach represents a shift in technical documentation strategy, suggesting AI systems will increasingly favor straightforward, accessible formats over complex documentation. This trend points to a future where technical interfaces are designed with both human and AI readers in mind.

“Clearly written instruction manuals in plain text are the future of APIs. I suspect this effect will be multiplied as AI models will start to use more accessible/less error prone methods of accomplishing their goals, something you can already see happening if you use AI a lot.” / X https://x.com/emollick/status/1902738826822308100

China-US AI gap narrows to three months
China has closed the AI development gap with the US to just three months in some areas, according to 01.AI CEO Lee Kai-fu. DeepSeek, a Chinese startup, demonstrated this progress by developing an AI reasoning model using fewer advanced chips at lower cost than Western competitors. Lee points to Chinese companies’ innovative approaches to algorithms and efficient chip usage as key factors, noting that US semiconductor sanctions have actually spurred innovation by forcing companies to work under constraints. DeepSeek has notably matched or potentially surpassed US capabilities in showing users AI reasoning processes, a feature developed but not released by OpenAI.

DeepSeek narrows China-US AI gap to three months, 01.AI founder Lee Kai-fu says | Reuters https://www.reuters.com/technology/artificial-intelligence/deepseek-narrows-china-us-ai-gap-three-months-01ai-founder-lee-kai-fu-says-2025-03-25/

DeepSeek Upgrades AI Model with Major Performance Gains
DeepSeek has released V3-0324 on Huggingface, featuring substantial improvements across multiple benchmarks. The model runs at over 20 tokens per second on a 512GB M3 Ultra using mlx-lm in 4-bit mode. Performance gains include a 19.8-point increase in AIME math scores and a 10-point improvement in LiveCodeBench. The update enhances front-end web development capabilities, producing more executable code and visually appealing web interfaces. Chinese language abilities have improved significantly with better writing quality and translation features. The model also fixes function calling issues from previous versions and can be run locally, making advanced AI capabilities more accessible to individual users.

“The new Deep Seek V3 0324 in 4-bit runs at > 20 toks/sec on a 512GB M3 Ultra with mlx-lm! https://x.com/awnihannun/status/1904177084609827054

AK on X: “DeepSeek just quietly dropped DeepSeek-V3-0324 on Huggingface https://t.co/6YFtRE5u3C” / X https://x.com/_akhaliq/status/1904154585242935516

“DeepSeek just quietly dropped DeepSeek-V3-0324 on Huggingface https://x.com/_akhaliq/status/1904154585242935516

Perplexity Proposes Plan to Rebuild TikTok with American Oversight
Perplexity released a proposal to transform TikTok by rebuilding it with US-based infrastructure and transparent algorithms. The AI answer engine company suggests developing an open-source recommendation system housed in American data centers, enhancing the platform with citation capabilities, and integrating Perplexity’s search functionality. Their plan includes implementing community verification features similar to Twitter, enriching videos with contextual information, and optimizing for user satisfaction rather than just engagement. Perplexity argues they occupy a unique position to rebuild TikTok without creating a monopoly, while addressing concerns about foreign influence on content algorithms.

Rebuilding TikTok in America https://www.perplexity.ai/hub/blog/rebuilding-tiktok-in-america

Lockheed Martin and Google Cloud Announce Collaboration to Advance Generative AI For National Security
Lockheed Martin and Google are integrating Google’s generative AI into Lockheed’s AI Factory ecosystem. The partnership aims to strengthen national security applications and will focus on intelligence analysis, decision-making, maintenance, and supply chain operation. It’s mostly a directional indicator of AI being growing part of national security.

Lockheed Martin and Google Cloud Announce Collaboration to Advance Generative AI For National Security https://www.prnewswire.com/news-releases/lockheed-martin-and-google-cloud-announce-collaboration-to-advance-generative-ai-for-national-security-302413514.html

Budget-Friendly Image Generator, Reve, Challenges Industry Leaders at Just a Penny Per Image
Reve Image 1.0, a relatively unknown AI image generator codenamed “Halfmoon,” is gaining attention for potentially outperforming established tools like Midjourney, Flux, and Ideogram at a fraction of the cost. The service offers 20 free generations daily and operates on a credit system, with 100 free credits at signup and additional batches priced at $5 for 500 generations, significantly less than competitors’ monthly subscriptions that range from $8 to $120. Despite limited information about its architecture and development team, Reve stands out for prompt accuracy, visual quality, and typography handling across various artistic styles. The platform also grants users ownership rights over generated images, unlike some competitors that claim ownership of content created with free plans.

Reve on X: “Halfmoon is Reve Image — and it’s the best image model in the world 🥇 (🔊) https://t.co/Zm1FzNQaFh” / X https://x.com/reveimage/status/1904211082870456824

“Holy crap. Reve is REALLY good. No surprise that it’s currently #1 in the Artificial Analysis text-to-image leaderboard — ahead of Recraft v3, Imagen v3 and FLUX 1.1 Pro. This team cooked, and it shows. https://x.com/bilawalsidhu/status/1904325105267683481

New Reve Image Generator Beats AI Art Heavyweights MidJourney and Flux at a Penny Per Image – Decrypt https://decrypt.co/311375/new-reve-image-generator-beats-ai-art-heavyweights-midjourney-and-flux-at-a-penny-per-image

“Excited to come out of stealth at @reveimage! Today’s text-to-image/video models, in contrast to LLMs, lack logic. Images seem plausible initially but fall apart under scrutiny: painting techniques don’t match, props don’t carry meaning, and compositions lack intention. (1/4) https://x.com/Taesung/status/1904220824435032528

OpenAI Nears Historic $40 Billion Funding Round Led by SoftBank
OpenAI is finalizing a $40 billion funding round led by SoftBank that would value the ChatGPT maker at $300 billion. The deal, structured in two tranches, includes SoftBank initially investing $7.5 billion with additional investments from Magnetar Capital (potentially contributing $1 billion), Coatue Management, Founders Fund, and Altimeter Capital Management. This valuation nearly doubles OpenAI’s previous $157 billion valuation from October 2024. While OpenAI projects revenue to reach $12.7 billion this year and $29.4 billion next year, the company doesn’t expect to become cash-flow positive until 2029 due to the significant costs of chips, data centers, and talent needed for AI development.

OpenAI Close to Finalizing $40 Billion SoftBank-Led Funding – Bloomberg https://www.bloomberg.com/news/articles/2025-03-26/openai-close-to-finalizing-its-40-billion-softbank-led-funding

Anthropic Releases Second Economic Index Report: Software Developers and Creatives Lead Usage
Anthropic’s second Economic Index report analyzes Claude 3.7 Sonnet usage patterns, revealing shifts in how people use AI. Since launching their newest model, Anthropic has observed increased usage for coding, education, science, and healthcare applications. The report shows the new “extended thinking” mode is primarily used for technical tasks by computer scientists, software developers, and digital creatives. The data also breaks down AI usage patterns between augmentation (where humans work alongside AI) versus automation (where AI completes tasks with minimal intervention). Tasks associated with copywriters show high collaboration rates, while translation tasks tend toward automation. Anthropic has also created a detailed taxonomy of 630 specific use categories to better understand how people actually apply AI in practice. All datasets are available for researchers to download and analyze further.

“Last month we launched our Anthropic Economic Index, to help track the effect of AI on labor markets and the economy. Today, we’re releasing the second research report from the Index, and sharing several more datasets based on anonymized Claude usage data. https://x.com/AnthropicAI/status/1905381798676197819

Chinese AI Agent Manus Seeks $500M Valuation as US Interest in Its Products Grows
Butterfly Effect, the Chinese startup behind AI agent Manus, is seeking tens of millions in funding from US venture capitalists at a $500 million valuation, five times its previous valuation. Manus, which can execute complex tasks like travel booking and stock analysis, has gained popularity on X despite US-China tech tensions. Unlike companies developing foundational AI models, Butterfly Effect focuses on applications, potentially avoiding US investment restrictions. Manus spends about $2 per task using Anthropic’s Claude model and reportedly exceeded $1 million in Claude expenses within two weeks of launch. With 2.6 million users waiting for access and plans for a Tokyo office, the funding would address server capacity limitations and support global expansion. Butterfly Effect has also partnered with Alibaba to use domestic AI infrastructure in China.

Chinese Startup Behind Manus AI Agent Seeks $500 Million Valuation — The Information https://www.theinformation.com/articles/chinese-startup-behind-manus-ai-agent-seeks-500-million-valuation

Browser Use Raises $17M to Help AI Agents Navigate Websites
Browser Use, a startup making websites more “readable” for AI agents, secured $17 million in seed funding led by Felicis with participation from Paul Graham, A Capital, and Nexus Venture Partners. Founded by Magnus Müller and Gregor Žunič through ETH Zurich’s accelerator, the company converts website elements into text formats that AI systems can better interpret. This approach avoids the limitations of vision-based navigation that relies on screenshots, allowing agents to complete tasks more reliably and at lower cost. The tool has gained significant traction among developers—including over 20 companies in Y Combinator’s current batch—and caught investor attention after being featured in the viral Manus tool from Chinese startup Butterfly Effect. Felicis investor Astasia Myers sees web AI agents as “the next frontier” for automating human tasks and creating a bridge between text-focused AI models and the dynamic digital landscape.

Browser Use, the tool making it easier for AI ‘agents’ to navigate websites, raises $17M | TechCrunch https://techcrunch.com/2025/03/23/browser-use-the-tool-making-it-easier-for-ai-agents-to-navigate-websites-raises-17m/

AI Detects Endometrial Cancer With 99% Accuracy
Researchers have developed an AI model that can identify endometrial cancer in tissue samples with 99.26% accuracy, significantly outperforming current automated diagnosis methods (which achieve around 80% accuracy). The model, called ECgMPL, examines microscopic tissue images by enhancing image quality, identifying key areas, and analyzing the tissue structure. Created by an international team from Bangladesh, Australia, and Canada, the system also shows strong results with other cancers, detecting colorectal cancer with 98.57% accuracy, breast cancer with 98.20% accuracy, and oral cancer with 97.34% accuracy. Researchers suggest this technology could become an essential diagnostic aid for physicians detecting multiple cancer types.

AI diagnoses major cancer with near perfect accuracy | Charles Darwin University https://www.cdu.edu.au/news/ai-diagnoses-major-cancer-near-perfect-accuracy

5 AI Visuals and Charts: Week Ending March 28, 2025

OpenAI GPT image teaser example “the pros and cons https://x.com/ajabri/status/1904599427366739975

“Sure, you could use an annotation tool to create bounding boxes of objects in images for you… or you can ask a multimodal AI to do it freehand. https://x.com/emollick/status/1904028116063822141

“The Screenshot: a fake screenshot generated by ChatGPT 4o of a Wikipedia article about the screenshot itself, with a copy of the screenshot in the article https://x.com/goodside/status/1904743355147235834

“I’ve had access to the new GPT-4o image generator for a bit: “now i need you to perfectly illustrate the moment that Elvis met Napolean at waterloo” “make it photorealistic” “napoleon is wearing a rubber duck on his head elvis’s pants have the ideal gas law printed on them” https://x.com/emollick/status/1904608706970398841

“The force is strong with this one. Pika’s new feature would make 12 year old me very jealous. Just type what you want to edit existing video — no after effects required. https://x.com/bilawalsidhu/status/1902873480170967383