AI AVATARS CAN NOW SPEAK FOR UP TO 5 MINUTES IN A SINGLE GENERATION OmniHuman 2.0 turns a character photo into a video with speech, facial expressions, gestures, and full-body movement. Upload an audio track, describe the camera movements and actions in text, and the model generates the performance. Here's what stands out: 5-minute videos: up from 60 seconds, without stitching separate clips together. Dialogue and music: support for more expressive content formats. Lower generation costs: reportedly down from $0.12 to $0.07 per second. At $0.07 per second, a five-minute video costs around $21. That opens up practical use cases for virtual presenters, advertising creatives, educational content, and branded characters that need to speak regularly. The bigger opportunity is content production at scale. Instead of filming a person every time you need a new video, you can build a repeatable workflow around a digital character, a script, and an audio track. The real test is whether the output stays consistent and natural across the full five minutes. https://t.co/PW8ib5olVc

BLACK FOREST LABS JUST MADE 4K AI IMAGE GENERATION POSSIBLE ON 8GB GPUs @bfl_ai released FLUX 3 Image Klein, a 5.6B-parameter open-weight model that brings advanced image generation and editing to consumer hardware. Instead of depending entirely on cloud APIs, developers can now run a capable image model locally through ComfyUI or Diffusers. Here's what it offers: 5.6B parameters Native 4K image generation Editing with up to 10 reference images Bounding-box controls for precise layouts Base and distilled models for different workloads INT8 support for GPUs with as little as 8GB VRAM According to Black Forest Labs, the distilled model generates a 1MP image in under 9 seconds on an RTX 4070. But the bigger opportunity isn't another speed benchmark. It's what happens when marketing teams can build their own image production infrastructure. Imagine a creative pipeline where agents generate dozens of ad variations, adapt existing assets to different formats, and maintain consistent visual references without sending every request to a separate SaaS platform. The model can run locally, its settings can be standardized, and repetitive production tasks can be automated through custom workflows. There's an important limitation. Open weights don't mean unlimited commercial freedom. The FLUX Kommunity License allows free commercial use for businesses with annual revenue below $5M, while larger organizations need additional licensing. Local inference also comes with hardware, maintenance, and setup costs. But the direction is becoming clear. As image models become smaller and more accessible, the ability to generate visuals stops being the main competitive advantage. What matters more is the system around them: creative direction, brand consistency, proprietary assets, and distribution. AI is making visual production infrastructure increasingly accessible. The teams that know what to create and why will benefit the most. research ↓

META JUST RELEASED AN 816B OPEN-WEIGHT AI MODEL THAT RUNS ON TWO MAC STUDIOS @Meta has released Muse Spark Open, its new AI model built for complex coding, autonomous agents, and long-running engineering tasks. The unusual part is that developers can download the weights and run a model of this scale on local hardware instead of relying entirely on cloud APIs. Here's what Meta packed into it: 816B total parameters 42B active parameters per token 1M-token context window Native text, image, video, and audio input Support for AMD, Apple, Intel, and NVIDIA hardware The architecture uses Mixture of Experts. Instead of activating all 816 billion parameters for every token, the model uses roughly 42 billion, reducing the computation required during inference. That makes its desktop deployment especially interesting. Meta worked with Apple to optimize Muse Spark Open for two M5 Ultra Mac Studios using the MLX framework. Together, the machines provide up to 1TB of unified memory. The company also demonstrated actual agentic workflows running locally: Fixing bugs and implementing features in an iOS application Completing software builds through Muse Code Controlling DaVinci Resolve to analyze footage Identifying strong video moments and creating vertical crops The model is available through Hugging Face and API providers including Groq, Together AI, Fireworks, and OpenRouter. But there's an important distinction. Open-weight doesn't mean unrestricted open source. Meta's Muse Model License includes commercial limitations, and running an 816B model locally still requires expensive hardware. What interests me most is what this could mean for AI-native operations. Marketing teams are increasingly building agents that process customer conversations, analyze competitors, classify content, and coordinate production workflows. Today, many of those systems depend on external model providers for every step. More capable open-weight models create another option: running sensitive research, internal knowledge processing, and specialized agents on infrastructure the organization controls. That doesn't automatically make local AI cheaper or better. Deployment costs, inference speed, maintenance, and model quality still determine whether it makes sense. But the competitive landscape is changing. Access to powerful models is gradually becoming less dependent on a handful of closed API providers. And as model access becomes more flexible, the advantage shifts toward the teams that can build reliable workflows, connect proprietary data, and turn AI capabilities into useful systems.

GPT-5.6 SOL JUST CHANGED THE ECONOMICS OF CYBERSECURITY Researchers used AI to help uncover a critical Linux vulnerability that could allow a virtual machine to access its host. The reported API cost was under $700 The interesting part isn't just the vulnerability itself. Human researchers still directed and verified the work, but AI made a highly complex investigation dramatically cheaper If discovering critical vulnerabilities becomes this affordable, the advantage shifts to whoever can find, validate, and patch them faster AI is lowering the cost of both attacking and defending infrastructure via @entropyIO

META IS BUILDING THE NEXT DISTRIBUTION LAYER FOR AI AGENTS @Meta and Sierra are working with Shopify, @stripe and Walmart on a protocol that lets AI agents authenticate and interact with businesses on behalf of customers. The obvious story is security. Agents need permission to access accounts, check orders and perform actions without being treated as bots. But I think the bigger implication is for marketing. For decades, companies optimized websites to convert human visitors. Now they may need to serve AI agents that research products, compare offers and make decisions before a person ever opens the website. That changes the buying journey. Product data, pricing, availability, reputation and machine-readable information could become just as important as the landing page itself. And the competitive advantage shifts toward being discoverable, understandable and accessible to machines. The protocol is still early, but the direction is worth watching. Brands spent years optimizing for search engines. The next challenge is making sure AI agents can actually choose and interact with them.

AI JUST MADE BUILDING AN ADOBE ALTERNATIVE A LOT CHEAPER One developer reportedly used Claude Opus 5.5 and Rust to build seven open-source creative apps covering Photoshop, Illustrator, Premiere Pro, After Effects and more. They're not production-ready replacements yet. Professional compatibility, rendering quality and reliability still matter. But the bigger story isn't Adobe losing seven product categories overnight. It's what happens when the cost of building complex software drops dramatically. For years, established software companies benefited from the enormous engineering investment required to compete with them. AI is starting to weaken that barrier. The real competitive advantage may shift from who can afford to build the software to who can make it reliable, earn user trust and build the strongest ecosystem. That's a much more interesting threat to Adobe than another free Photoshop clone.

Prediction markets have a vanity metric problem: trading volume. Volume looks like traction, but it can spike around major events and count the same capital repeatedly as traders rotate positions. Open interest tells a much cleaner story. As of May 1, tracked prediction markets held roughly $1.11B in OI. Kalshi: ~$630.7M Polymarket: ~$449.9M Together, they accounted for almost 98%. Everyone else was below $15M. That changes how I read the category. This isn't ten platforms competing on equal footing. It's two liquidity centers followed by a long tail of increasingly specialized products. @Kalshi is winning with regulated rails and brokerage-like UX. @Polymarket owns the crypto-native position. Smaller players are differentiating around specific environments: exchange-style crypto markets, macro contracts, or prediction markets embedded directly into existing trading venues. The broader lesson is simple: activity is not the same as depth. You can subsidize volume and make a dashboard look active. You cannot fake whether serious capital can enter and exit without moving the market. I'd rank prediction markets by OI, depth, spreads and resolution quality before headline volume.

Perp DEXs in 2026 don't have a traffic problem. They have a retention problem. Getting 50,000 wallets through KOLs, points and an airdrop campaign is relatively easy. Getting a trader to deposit, open a position, come back next week and eventually make your venue part of their trading stack is much harder. That's why @entropyIO is interesting to watch. Instead of relying only on speculative future rewards, Entropy is building incentives around behavior that can compound: trading rebates, referrals, partner distribution and rewards paid in $HYPE using deployer-level fees. It recently started distributing those rebates to traders and referrers, while also expanding distribution through channels like Trust Wallet. And there is already meaningful activity to retain. DefiLlama currently shows roughly $895M in 30-day perp volume for Entropy. The important metric now isn't how many wallets a campaign can acquire. It's funded accounts → first trade → second week → Day 30 retention → LTV. Points can rent volume. A real growth system makes coming back economically rational even after the campaign ends.

THE DEAD ACCOUNT IN THE FUNNEL IS OFTEN THE FOUNDER’S Every audit looks the same. The brand posts every day. The founder’s personal account froze in April. A pin from last season’s launch. For a product that needs users and partners, that gap is expensive. Nobody owns closing it. Teams file the channel under “personal” or “when I have an hour.” In Web3 the founder’s voice is the highest-trust asset the company already has. Most teams leave it off. THIS IS NOT ABOUT CHARISMA Edelman 2025: scientists 77%, teachers 75%, CEOs in general 53%. People trust a person who knows the subject. Not the org chart. In crypto, attention also has a price. Mindshare pools around a face and a point of view. A faceless brand fights for that attention without its strongest asset in play. The brand says what shipped. The founder says why it matters and whether the person building it is worth the bet. WHERE INBOUND ACTUALLY LANDS Partnerships, strong hires, a fund intro, an exchange conversation rarely arrive in the brand inbox. They land in the DMs of the person whose take someone finished reading. The personal account is top of funnel. Not a Friday thought. HOW NOT TO KILL THE CHANNEL Brand: announcements, product, facts. Founder: where the market is wrong, how you build, the occasional loss. The same press release from two accounts is dead weight. Do not count followers. Count replies from the right people. Threads that open. Calls. Deals you can tie to a post. Five thousand new followers and zero conversations is a failed month. Half the channel is not your own timeline. It is a reply under the accounts your buyer already reads. The channel cannot depend on a free evening. That is the week it dies, and the week you needed it. LinkedIn is a separate room. Trust from Crypto Twitter does not move with you. An investor searches the name. An empty profile or “gm, new airdrop” is not credibility there. It is a signal you are talking to the wrong hall. A company page gets a fraction of the reach a person gets. A silent founder makes every later touch more expensive. A live voice makes them cheaper.

YOUR NEXT CUSTOMER MIGHT BE A BOT For a long time marketing had one job: get a human on the page. The tab is often opened by something else now. A browser agent reads, compares, and more and more finishes the action for a person who never saw the site. In April 2026 about 71% of agent traffic came from browser agents. Perplexity’s Comet and OpenAI’s Atlas accounted for about 70% of that layer. Commerce was the second destination, about 38% of the routes. And they buy. Adobe: visits to US retail sites from AI sources grew 393% year over year in Q1 2026. In March those visits converted about 42% better than the rest. A year earlier it was the opposite: 38% worse. Salesforce: in the 2025 holiday season agents closed about 20% of global orders, roughly $262 billion. WHAT THE AGENT REWARDS It does not feel the founder story. It takes facts: what this is, the price, the terms, whether the action can finish without getting stuck. Most sites, Web3 especially, are hostile to that. Metaphors. Motion. The point on the third click. A human will wait a second. The agent leaves for a page it could actually read. The audience split. A person still needs a story and proof. A machine needs a page it can parse and buy from. That is already a channel, the way a Google ranking used to be. WHERE CRYPTO HAS A HEAD START An agent needs rails built for software, not a button made for a thumb. Stables, programmable wallets, and onchain settlement sit closer to that than a card form. Only about 3% of agent sessions reach checkout today. That window will not stay open. Keep the facts in text: product, price, terms. Split agent traffic in analytics. Close the one main action in a short path, without extra modals and wallet detours. For twenty years search meant: they found you and sent a person. The next layer: the agent understood the page and either took you or did not. In a fraction of a second. No loyalty. A small team with a clear, readable page can beat a beautiful site the machine could not parse.