Poonam Soni

JUST IN: Anthropic is meeting White House officials today to get Claude Fable 5 and Mythos 5 back online.

@ai_for_success Amazing result

One of the reasons I love using Google Gemini Omni Flash is that it matches your voice. While creating an avatar, it clones your voice and it’s actually really good.

JUST IN: a Stanford AI lab just took #1 on both the text-to-speech AND speech-to-text leaderboards. same week. nobody has ever done that. it’s called Cartesia. and they just shipped two models at once. - Sonic 3.5: text-to-speech - Ink 2: speech-to-text

PS. I share more in-depth insights on “AI tools” and their future developments here. Join over 35K+ readers in the AI Toast community for free: aitoast.beehiiv.com

Source: cnbc.com/2026/06/15/ant…

JUST IN: Anthropic is meeting White House officials today to get Claude Fable 5 and Mythos 5 back online.

@DanutPralea Yeah..

@CodeByPoonam the smartest model on earth needs a permission slip from the white house. what a timeline

@krandiash 82ms to first audio is genuinely insane. ElevenLabs Turbo is sitting at 250ms+. this is not close.

We released Sonic-3.5 and Ink-2, the #1 streaming models for text to speech and speech to text you can use in your voice agents today. New architectures enable new frontiers for speed and quality. We're now the only provider to have #1 models for both speaking and listening.

@LaionardoArt cnbc.com/2026/06/15/ant…

@CodeByPoonam Source?

Elon Musk just said SpaceX will hit $1 trillion in revenue by 2030. Wall Street says $330 billion. someone is very wrong. Musk's exact words: "I think SpaceX might be able to reach approximately $1T revenue in 2030." then doubled down: "I would be surprised if revenue is not greater than $1T in 2031." the math: → SpaceX revenue 2025: $18.7 billion → Musk's 2030 target: $1 trillion → that's a 53x jump in 5 years what's fueling the bull case: → Starlink: broadband for the entire planet. barely started. → Starship: 10x cheaper launches. entirely new markets. → Google paying $920M/month for SpaceX compute. → Anthropic paying $1.25B/month. → AI infrastructure in space: a category that didn't exist 2 years ago. what Wall Street thinks: → Morgan Stanley: $330 billion → Goldman Sachs: well short of $1T → ARK Invest: $300-400 billion the gap: $670 billion. they said reusable rockets were impossible. they said Starlink would never work. they said Tesla would go bankrupt. every time Wall Street doubted him, he proved them wrong.

@heyrobinai Minimax is good

CHINA JUST DROPPED AN OPEN-WEIGHTS MODEL THAT COMPETES WITH CLAUDE, GPT, AND GEMINI. MiniMax M3 goes head-to-head with Claude Opus 4.7, GPT-5.5, and Gemini 3.1 Pro. And I tested it myself. It's the first open-weights model to combine three things that were closed-source-only until last week: → Frontier coding & agentic performance: 59.0% SWE-Bench Pro (ahead of GPT-5.5's 58.6%), 66.0% Terminal Bench 2.1, 74.2% MCP Atlas → 1M token context window via MiniMax Sparse Attention, at roughly 1/20th the per-token compute of their previous gen at full context → Natively multimodal from step zero: image + video input, can even operate a desktop The weights are now live on Hugging Face. You can run this thing locally. But benchmarks are one thing. I wanted to see what it does with a real-world task. THE TEST One single prompt: build a complete, client-ready company profile website for a fictional tech consultancy called "AstraCore Solutions." Not a toy landing page. The requirements: > 11 full sections: sticky header, hero, company overview with metrics, 6 service cards, 4-step process, 3 case studies, why-us, testimonials, FAQ, contact form, complete footer > React + Tailwind CSS > Zero lorem ipsum. All real, persuasive copy > Strict design rules: light theme, tech blue + warm accent, max 8px border radius, hover states, lucide-react icons, polished mobile layout > SEO metadata, clean component structure > And the kicker: after building, it has to RUN the app and fix its own build/lint errors This tests exactly what matters in practice: long instruction-following, sustained content generation, design taste, and closing the agentic loop by verifying its own work instead of just dumping code. WHY THIS MATTERS Pricing is $0.30/$1.20 per 1M input/output tokens up to 512K context. A fraction of what the closed frontier costs. If an open-weights model can deliver client-ready output at this price, and you can self-host it for compliance-heavy environments, the calculus for a lot of dev teams changes overnight. Caveats worth knowing: the benchmark numbers are company-reported, and their comparison baseline is Opus 4.7, not the newer 4.8. Independent testing (Artificial Analysis) puts it at 55 on the Intelligence Index, making it the leading open-weights model, just ahead of Kimi K2.6. Results of my test below 👇

@mhdfaran This is awesome

Making 3D assets just got stupidly easy. Tripo turns a sentence or a single image into a full 3d model, ready for blender, games, printing, whatever. Idea → image → 3D asset. Full workflow below 🧵

@jerrod_lew Looks cool

Progress on my spinning top battle game. Vibe-coding this with Claude Code, Opus 4.8 and using assets created with the Magnific MCP. Main improvements are the 3D models with the Magnific MCP. 1. 4 top varieties converted to 3D model. 2. Batle stadium 3D model. Also added new attacks, collisions, health bars, jump and dash buttons.

@RrichPRMR Exctly 🥲

@CodeByPoonam the future arrived in a box smaller than my keyboard

@aaliya_va yes

@CodeByPoonam Good point. Companies that keep learning will grow faster.

damn.. Satya Nadella just said something every CEO needs to read right now. "You can offload a task or a job. But you can never offload your learning." He introduced two terms that reframe the entire AI conversation. → Human capital: the knowledge, judgment, relationships, and pattern recognition of your people. → Token capital: the AI capability your company builds and owns. The counterintuitive part: human capital doesn't become less valuable as AI grows. It becomes more valuable. Without human direction, you just have compute running in circles. The real opportunity isn't picking the best model. It's building a learning loop where both compound together. Private evals that measure whether AI is improving against your outcomes, not external benchmarks. Reinforcement learning on real traces from inside your organization. A knowledge base that makes institutional memory queryable and reusable. This loop becomes the new IP of the firm. And unlike most assets, it compounds. Every improved workflow generates better training signal. Every better signal accelerates the accumulation of knowledge unique to your company. the warning nobody is talking about: Satya compared AI commoditization to globalization. a small number of models eating everything. entire industries finding their knowledge commoditized out from underneath them. "The political economy will simply not tolerate it." his conclusion: a frontier without an ecosystem is not stable. the companies that build their own learning loop early will have an advantage that is hard to replicate. regardless of any new model capability.

@icreatelife So cute

It's Sunday but who else is making an AI animation?

@IamEmily2050 @Kling_ai Amazing

Kling 4K

@amit_esquire Thank you :)

@CodeByPoonam I’ve subscribed. Looking forward to reading and learning more about AI.

Google Omni video editing is too powerful you can now edit videos through plain conversation.