imagine a robot trying to pick up an egg. it can see the egg, figure out where to grab it, and plan the movement. but how does it know how much force to apply without crushing it? that's something we humans do without even thinking. Silica is trying to solve this with actuators that are 50x more sensitive, at an affordable price. Such a cool idea, and I feel this angle of robotics should be explored.

building with AI models right now means jumping from one platform to another. one for image generation, another for video, another for audio, and then managing separate API keys, billing, and integrations for each. @sunra_ai just launched a unified API that brings image, video, audio, and 3D models under one roof. one API key, pay-as-you-go pricing, and access to models at their makers' own API prices without markup.

i think we're missing something with all these AI shopping agents. half the time i don't even know what i want to buy. i just browse until something catches my eye lol @daletlabs is doing something interesting with scene. it generates realistic images based on your taste, and you can actually buy the products you see, across different brands, in one checkout. probably one of the more practical applications of generative media i've seen recently.

this might be the strongest case for agent swarms yet. @axiommath just formalized one of the largest proofs in mathematics. ~15,000 pages of work, verified in 24 days. agents coordinated through GitHub, reviewed each other's work, and even found a decades old flaw. i think this is where software engineering is heading. code we can actually verify.

came across @_MonoDesk today, which handles everything from proposals to invoices in one place. say a client asks for an extra feature. its agent can check the agreed scope, flag the additional work, and draft a response. and it's actually 100% free and has MCP integration!

human intelligence has always been distributed. most of what we've achieved comes from people with different capabilities working together. but with AI, we're mostly trying to build superintelligence by scaling individual models. i came across @dcog, which just came out of stealth with backing from Verdict Capital. they're exploring Topological Super Intelligence (TSI), where smaller, specialized models collaborate through learned communication layers. i find the idea that communication between models could matter as much as model size quite interesting.

one of the rarest things in AI today is data that models haven't already seen. a lot of valuable data is sitting inside private companies, completely inaccessible to the AI labs that could actually use it for training. i recently came across @datavendor_ai, which is building a marketplace for exactly this. instead of spending weeks finding vendors and negotiating access, AI labs can simply post what they need, and vendors can come to them with matching datasets. i feel this is a pretty interesting direction for how AI training data could be sourced and traded in the future.

i recently found out how much money sales teams lose even after someone has already shown interest. imagine spending thousands on acquiring leads, only to lose them because nobody followed up fast enough. people start booking, drop off halfway, and by the time someone reaches out, they've already moved on. that's why i found @iclosedsaas's new iClosed Unified Inbox interesting. it captures leads even if they never finish booking, then follows up in under 5 minutes on the channel they're most likely to answer.

brands are probably missing most of the conversations happening about them on the internet. someone could be talking about your product in a viral video without ever tagging you, and you'd never know. that's what i find interesting about @nectarsocialai. it uses multimodal AI to catch these conversations, discover creators, and deploy agents to engage in the brand's voice. this is how AI agents in marketing get genuinely interesting, generating actual growth :)

Used Treg right now to find some leads for AI startup founders. It actually found so many results in such a short time and only cost me a few cents. Gonna be a power user now :)

I've always been interested in pushing local inference without compromising model quality, and Recursive's Flint looks really promising. 40B MoE with just ~2B active parameters, scoring 89.7 on SWE-bench Pro v2 and 89.2 on IFBench. What's more interesting is that their Eureka Machine helped optimize the architecture, low-bit training, and inference kernels. Really interesting direction for automated AI research.

New in lattedit: a citation library that follows you across papers. Cite from the toolbar and the entry lands in your .bib. Plus author kits for NeurIPS, ICML, ICLR, ACL, CVPR and more, and a bug report button. Free → lattedit.xyz

AI has made it easier than ever to create images, and somehow managed to make design worse than ever. Every AI-generated image looks painfully similar now. Same aesthetics, same generic layouts, and honestly, it's getting harder to find anything that actually looks original. I made a fashion week poster with Vyb by @cocoailabs. It uses actual brand assets to create layered, editable designs instead of another generic AI image. I think this is the direction AI-generated design should have taken from the beginning. More control over what you're creating, rather than just hoping the model gets it right.

Something I've found really interesting while looking into LLM inference is how autoregressive decoding is often limited by memory bandwidth rather than compute. More FLOPs don't necessarily mean more tokens/sec. That's why the @general_compute x @positron_ai partnership caught my attention. GPUs handle compute-heavy prefill, while Positron's memory-first chips target bandwidth-heavy decoding. Atlas becomes available in Q1 2027, with Titan to follow. I think heterogeneous inference is where AI infrastructure is heading.

Never thought I'd find a robot dancing this adorable, but here we are. This is Voxel, a little robot from @roboticscorp, and I genuinely love the direction they are taking. What I find really interesting is how much they're focusing on personality and design rather than making it feel very robotic. They even worked with Pixar on this. I can actually see myself wanting one right now.

Something I find really interesting about @skyfallai is how they're bringing world models into actual engineering workflows. Their new model, SKY-001, can simulate designs, break complex objects down into individual components, and then reconstruct them in 3D. They're reporting 100x faster inference and 150x lower costs than GPT Astra for CAD tasks, which is pretty impressive. And looking at their roadmap, CAD seems to be just the beginning. They're exploring everything from PCB design and robotics to enterprise workflows. Really curious to see how far @spisallyouneed and the team can take this, especially at this pace.

Your model is already good, but what if it's working with the wrong context? You can give a model the best reasoning capabilities in the world, but if the information it's working with is wrong, the answer will be wrong too. That's why I find what @mynameisyahia is building with Context really interesting. I feel like the next big step towards better AI agents isn't just improving models, but building the infrastructure that makes the internet more accessible to them. Better context means better answers.

I've gotten pretty good at figuring out SEO, but AEO is a completely different game. I've struggled with it myself, and the fact that people are paying thousands of dollars just to figure out AEO says a lot about how difficult it still is. Really cool to see Treg launch something that tackles this. I'll definitely be using it for my own products. And the crazy part? It costs just 1 cent.

One thing I really wanted to fix with lattedit was how painful it is to actually work on LaTeX projects with other people. So we built real-time collaboration, inline comments, and an AI agent powered by Sarvam 105B that can directly edit your LaTeX. You can now also convert PDFs back into editable LaTeX while preserving the layout, import your Overleaf projects, and start with ready-made templates. And yes, everything is free. lattedit.xyz