Jason Ederle - Stealth AI Startup | LinkedIn
Jason Ederle
Experience & Education
-
Stealth AI Startup
* * ###
* * ###
** undefined
Patents
Issued December 28, 2016 US US20170223005A1
Technique for local device authentication.
Issued December 15, 2014 US US20160173436A1
Patent for lightweight event planning.
View Jason’s full profile
- See who you know in common
- Get introduced
- Contact Jason directly
Other similar profiles
Ahmed Aly
Meta
14K followers
Greater Seattle Area
Baiyang (Keen) Wang
11K followers
Newark, CA
Denis Brockus
Amazon Web Services (AWS)
6K followers
United States
Brian Yang
OpenAI
11K followers
San Francisco, CA
Navdeep Dahiya
Meta
7K followers
United States
Haoran Liu
Tencent
3K followers
San Jose, CA
Vimal Kumar Selvam
Microsoft
13K followers
Renton, WA
Amit Patankar
Blue Voice Inc.
6K followers
San Francisco Bay Area
Zhia Chong
Meta
7K followers
Greater Seattle Area
Bhupendra Kastore
9K followers
New York, NY
Explore more posts
-
Arian Barvarz University of West London • 537 followers 𝗩𝗥 𝗶𝘀 𝗵𝗮𝘃𝗶𝗻𝗴 𝗮 𝗿𝗼𝘂𝗴𝗵 𝗺𝗼𝗺𝗲𝗻𝘁. 𝗔𝗻𝗱 𝗮𝘀 𝗮 𝗳𝗮𝗻 𝗼𝗳 𝘁𝗵𝗲 𝗺𝗲𝗱𝗶𝘂𝗺… 𝘁𝗵𝗮𝘁 𝗵𝘂𝗿𝘁𝘀. With Meta Reality Labs cutting back and pivoting harder toward AI, AR glasses, and wearables, a bunch of genuinely great VR studios have been hit in the crossfire like: Armature, Sanzaru Games, Twisted Pixel and Camouflaj (heavily reduced) These teams helped define what good VR actually feels like. From Resident Evil 4 VR to Asgard’s Wrath, to some of the most polished, thoughtful VR experiences out there this talent shaped the medium. And losing them isn’t just sad for the people (𝘄𝗵𝗶𝗰𝗵 𝗶𝘀 𝘁𝗵𝗲 𝗺𝗼𝘀𝘁 𝗶𝗺𝗽𝗼𝗿𝘁𝗮𝗻𝘁 𝗽𝗮𝗿𝘁)… It’s also a long-term hit to VR itself. Here’s the thing people forget: Great platforms are built on great content. And great content comes from teams who’ve spent years learning what works in VR comfort, interaction, presence, pacing, embodiment. That kind of knowledge doesn’t magically reappear overnight. If: • A new Half-Life VR game (Alyx) • A solid, affordable high quality headset like Quest 3 • And years of platform investment Still wasn’t enough to kick VR into true mainstream adoption, then yeah, we probably have to be honest with ourselves. VR isn’t mainstream, not yet anyways but that doesn’t mean it’s dead. It means we’re still early, still awkward, still figuring out what this medium actually wants to be and can be. From Meta’s perspective, burning billions forever was never going to be sustainable. The pivot is rational. But from a creator and fan perspective? The layoff? It still stings a lot! And make no mistake, the people leaving today are the ones whose work we’ll miss years from now. • VR will keep evolving. The tech will get smaller, lighter, cheaper, and better. • AR, wearables, mixed reality, it’s all part of the same long road. • But right now? This feels like a bump in the road for VR as a games platform. Really rough time for a lot of talented people who helped build something special. Wishing everyone affected nothing but the best, the industry is better because of the work you did. Even if the medium isn’t done growing up yet. #meta #oculus #gamedev #gamesindustry #vrgames #virtualreality #gamecareers #gameproduction #devcommunity #VR #AR 10 1 Comment
-
Ilia Ovsiannikov Remake.ai • 1K followers Founder Update: The Drawbacks of Using Claude.ai, Part 1/2 (or: Why I Love Claude No Matter What) TL;DR for busy professionals: - Claude codes like a brilliant junior developer — and here’s how to compensate for that. - I’ve managed to break Claude (and figured out how to fix it). - My real pain is Claude’s presentations and artifacts (and why). - Sometimes Claude doesn’t know something — and that’s OK (easy fix). - I still love Claude — she gives me a 48× productivity multiplier. In a previous post, I promised to write about the drawbacks of using Claude.ai. I’ve been delaying it — because, honestly, I’ve struggled to find any. I’ve been using Claude for about three months, and I meant it when I wrote that Claude has been my de-facto cofounder while I search for a human one. - Someone to bounce ideas with. - Erudite. Available 24/7. Eager to help. - All for $20 per month. Claude has built the entire MVP prototype for my startup Remake.ai — and I could not be happier with the results. But, yes — there are drawbacks. Let’s get into them. 👧1. Claude codes like a brilliant, yet junior developer Claude’s first-pass implementations consistently look like the work of a brilliant but junior dev. Here’s my workflow: I prompt Claude to implement a feature — without over-specifying how. Claude builds it the way she prefers. Then I test it. Usually, I ask for refactors: break large files apart, adjust class structure, improve reusability. I prefer it this way. Because I’m developing a non-standard system, I often don’t know the final structure yet. Claude’s first pass helps me see possibilities I wouldn’t have thought of. Claude needs senior-level supervision — someone who can architect, review, and guide. Without that, the codebase will quickly turn messy. I feel that consistently. When I combine Claude’s energy with my architectural judgment, the result is stunning: a year of work done in a month, while personally coding maybe 25% of it. That’s my 48× productivity multiplier. 🤕2. How I broke Claude (and how to fix it) Sometimes, Claude just... gets the flu. Performance drops, attention to detail slips. Once I noticed the async keyword missing in front of some Python functions. I prompted Claude to fix that - and offhandedly referred to those functions as “non-async,” A couple of prompts later, I noticed Claude started forgetting to prepend async during refactoring throughout the code base. I realized the issue was probably me, my inaccurate prompting. The fix: -💡 Start a new chat to reset context. Claude recovers instantly. Lesson learned — words matter more than I thought. 1 1 Comment
-
Ben Singer PICO XR • 7K followers if you're focused on gaming, XR, or Gen Z / Gen A, this is a spot-on post about the way - and why - they socialize online. great quick read. the point for so many users has never been to represent one's physical self in XR, and this is particularly true for younger people exploring their identity and agency. there's a whole corollary discussion about trust and safety to be had, but these ideas are critical to understanding that topic as well. #vr #genz #meta #roblox 16 3 Comments
-
Avi Bar-Zeev RealityPrime® • 8K followers Roblox is like a giant shopping mall full of people of all ages and a thousand different reasons for them the be there, some good, some bad. Would you let your 7 year old walk around a shopping mall alone or even with some same-age friends? What if the different kinds of shops only allowed people older than X years old to enter? For me, that wouldn't solve the core problem of the common hallways having a mix of everyone milling about and interacting. It's both the power of massive virtual worlds and the fundamental safety flaw. Not all people are good intentioned, or behave appropriately. Many are too young to know the difference, or old enough to know how to get away with it. The best way to make it safe IMO is to heed lessons learned by other platforms. Younger children can only interact with a safe-list of other children that the parents know and approve. The interactions themselves should be very limited for younger kids (not full text) and increasingly permissive for older ones as they gain independence and trust. This is not the same as age-gating the games. That's like having stores that only let in people of certain ages. AI moderation won't solve this. That's like having Robocop patrol the streets. Has Roblox done this yet? I'd love to know. 1
-
Nemanja Divjak I'm building ApexAPI… • 6K followers Guide Labs just open-sourced an LLM where every single output can be traced back to its training data. No more black boxes. No more "trust me bro" from your AI. Steerling-8B bakes interpretability directly into the architecture — built from the ground up, not bolted on after. The kicker? It hits 90% of frontier model performance with LESS training data. As someone building AI products daily, this is what I've been waiting for. Regulated industries, copyright concerns, trust issues — all solved at the architecture level. The era of "we don't know why the model said that" is ending. https://lnkd.in/dgcV3bwe #AI #LLM #Interpretability #OpenSource #MachineLearning 1
-
Charlie Fink Chapman University • 39K followers This week on The AI/XR Podcast, Caspar Thykier Thykier, CEO and co-founder of Zappar , has been building in XR since 2010. His take on how to survive 16 years in a space that has restructured itself every two to three years: "In XR, you've gotta be a bit cockroach-y. You've somehow gotta survive not getting squished." The companies that fell by the wayside had good technology. The differentiator for the ones that made it through wasn't the technology — it was the relentless focus on genuine utility. Not AR for a demo. AR for a real problem that someone with a real job needs solved. Caspar's other reflection: the industry hasn't always been great at explaining what AR actually does for a specific end user, which is why "AR can do everything" often landed as "I don't know what to do with it." The clarity is finally arriving. Watch the full conversation on the AI XR Podcast. #XR #AI #AIXRPodcast #Zappar #SpatialComputing 15 3 Comments
-
Alex Petrenko ZibraAi • 7K followers Zibra AI compression technology just cut training convergence time by ~12%. Recently, we were approached by a research lab to validate whether our solution could be applied across other volumetric data formats to accelerate model training - and the results were surprising. We’re now actively exploring this vertical expansion into simulation-heavy spatial AI domains, so if this resonates, feel free to reach out. Great post by Alexander Puchka on the “memory wall” challenge ⬇️ ⬇️ ⬇️ #ZibraAI #SpatialAI #AIInfrastructure #DataCompression 53 2 Comments
-
John Reagan Ulshe AI • 8K followers We’ve normalized software bloat. Modern operating systems require gigabytes of RAM and massive hardware acceleration just to render transparent windows. What happens when you throw out 30 years of legacy APIs and build the compositor from absolute bare-metal math? Today, we pushed a massive update to PolyMorphOS, our
#![no_std]Rust sovereign operating system. In the video below, you are looking at a 3D Holographic UI rendered entirely using Anisotropic Gaussian Splats. There are no bitmap textures. There is no GPU acceleration. It is running flawlessly at 60fps on a single CPU core (Core 0 / BSP) with just 64MB of emulated VRAM, while the other 3 cores are in a hardwareWait-for-SIPIsleep state. To the engineers at NVIDIA, AMD, Intel, Qualcomm, and Apple: the silicon you build is infinitely more capable than modern OS software allows it to be. By writing direct to DMA buffers and using pure integer math, we’ve bypassed the bloated rendering pipelines completely. But a sovereign UI needs a sovereign file system. In this release, we also deployed the MICT File System (MFS). Files in MFS aren't passive data, they are self-governing smart contracts. The kernel intercepts NVMe hardware reads, spins up a zero-allocation Virtual Machine on a 1-Kilobyte stack, and evaluates cryptographic hashes against compiled bytecode. If an AI agent lacks the correct context, the file physically rejects the OS. We are open-sourcing the blueprints to the factory. Stop renting APIs and start building sovereign intelligence. Code, whitepapers, and architectures available on GitHub (Free for research under BOKRLv2). 🔗 https://lnkd.in/gVS-3Kz3 #SoftwareEngineering #Rust #OperatingSystems #Semiconductors #ArtificialIntelligence #GaussianSplatting #BareMetal #NVIDIA #Intel #AMD 15 7 Comments -
Marc Alloul Groupe W inc • 4K followers I will be in San Jose, California March 16-19th attending NVIDIA GTC along Brandon Da Silva, ArenaX Labs, CEO. Looking forward to discuss about SAI, catch up with business partners and make new friends. (ping me separately if you are around). _ NVIDIA GTC (GPU Technology Conference) is a premier global AI conference focused on accelerating the future of AI, computing, and graphics. Held annually, it features keynotes from CEO Jensen Huang covering advancements in AI infrastructure, robotics, and simulation. The event is scheduled for March 16-19, 2026, in San Jose, California. _ ArenaX Labs Advancing Machine Learning together, is an AI technology company accelerating progress toward Artificial General Intelligence (AGI) by building platforms where learning agents evolve through competition and human interaction. SAI (https://lnkd.in/exb8YJGn) The Indispensable Evaluation Layer for the Robotics Economy Verifiable RL Benchmarking #NVIDIA #GTC2026 #AI #DeepLearning #CanadaTech 35
-
Mrunal Gawade HOLOFIL • 5K followers A tear down of the just launched Meta Ray ban display glasses version 1 ($800) showing its Geometrical waveguide and LCOS mini projector light engine, along with other important components such as battery (900 mWh or 248mAh capacity) and Snapdragon processor. Most important optical difference from other wearables is the the use of a Geometric waveguide than the traditionally used Diffraction waveguide in the lens. Most of the headsets and many glasses in the market so far (Magicleap, Hololens, Snap spectacles, etc.) all use diffraction waveguides. If you do not know what is a waveguide and the optics works, see the link in the comments, for educational understanding. As a background knowledge, the geometric waveguide uses tiny mirrors embedded in the lens sandwitched between the glass like material layers, to reflect the light thats totally internally reflected after projecting from the side projector, while the diffraction waveguide uses diffraction gratings to diffract the light and are made with mostly lithography machines (Nano-tech). Geometric waveguide offers the benefit of no color dispersion that is typically is seen in diffraction waveguides where the person looking at the glass from outside notices a rectangular color tint in the glass. Geometric waveguide also offers better color quality due to use of the mirrors in the materials than diffraction gratings. #optics #diffraction #light #meta #ar #display https://lnkd.in/etekyrvP 6 1 Comment
-
Navveen Balani Green Software Foundation • 13K followers ✨ From single prompt to Mini AI Apps - Google Labs has just launched Opal, an experimental new platform that empowers creators and developers to build powerful AI mini-apps using nothing more than natural language and visual editing. With Opal, you can: ✅ Chain prompts, AI models, and tools into workflows – no code required ✅ Visually edit flow logic or tweak steps on the fly ✅ Share your mini-apps instantly using just a Google account 🔍 What Can You Build with It? Whether you're: - Prototyping AI ideas - Building internal productivity tools - Crafting proof-of-concepts for clients 🚧 Opal is currently in public beta – not yet available globally. Opal offers a glimpse into a future where AI development is accessible to everyone — not just engineers. This is how AI will scale: not through code, but through creativity. 🔗 Learn more: https://lnkd.in/dfFUwcxT #opal #genai #google #GoogleLabs #AItools #NoCode 11
-
Montgomery Singman Radiance Strategic Solutions • 28K followers NVIDIA just put a universal beautifier on your AAA game. Gamers noticed. DLSS 5 was announced at GTC 2026 as Jensen Huang's "GPT moment for graphics." The demo footage applied generative AI on top of titles like Resident Evil Requiem and Hogwarts Legacy — reshaping character faces in real time, homogenizing skin tones, deepening wrinkles, smoothing features toward some algorithmic ideal. The community response was immediate. The announcement was ratio'd: 60,000 critical likes versus 40,000 approvals. Jensen's response? Gamers are "completely wrong." He may be technically correct about developer control. But this is still a misuse of AI. AI in games should be invisible infrastructure — faster rendering, smarter NPCs, more believable physics. The moment it starts making aesthetic decisions that belong to the artist, it has crossed a line. DLSS 5, as demonstrated, doesn't enhance an artist's vision. It replaces it with the model's. I've spent nearly four decades in this industry — writing code at Capcom and Electronic Arts, licensing titles into China, watching players in two very different cultures form deep bonds with game worlds. One thing holds true on both sides of the Pacific: players will forgive imperfect graphics. They will not forgive being told that what they loved wasn't good enough and needed to be fixed by a machine. Telling 60,000 people who just expressed that concern that they are "completely wrong" compounds the error. DLSS 5 launches in fall 2026. There is still time to reposition this — not as a universal beautifier, but as a tool that serves the artist rather than supplants them. That is the only version of this technology that belongs in games. Is this a turning point for how our industry thinks about AI's role in creative work? I think it should be. #Gaming #AI #GameDevelopment #NVIDIA #DLSS #ArtificialIntelligence #GameDesign #CreativeAI #VideoGames #GameIndustry 623 51 Comments
-
Henry Inskip Kratt Brothers Company Ltd. • 430 followers Few recommendations and learnings for those wanting to experiment with local private AI: 1. LM Studio - easy to use application that simplifies downloading and running inference on open-source models, works on Windows (CUDA, llama.cpp) and MacOS (MLX). There is also Ollama which is popular too. Choice is really personal preference here. 2. Hugging Face - get familiar with this site. Hugging Face is the central hub for exploring/download new models, datasets, as well as testing AI applications "Spaces" built by the community. 3. Small models are good now - there are some amazing massive 600B+ parameter models out there (DeepSeek/Mistral AI Large etc) but for a lot of use cases the small 7b-32b range of models work great. Shoutouts to the Qwen3 line of models as well as Google DeepMind's Gemma models and as of yesterday the new Mistral-3 models too. 3. Fine-tuning > massive API models. Benefits of running smaller models is that not only is your data local and private, but you can also fine-tune them on your own data and make them better than the massive frontier models for your specific use case. There are some great platforms and applications to help with fine-tuning and learning about the different methods (Supervised, Reinforcement Learning etc). Shoutout to Unsloth AI who are constantly putting out easy to use collab notebooks/detailed instructions. 4. Linux/Windows vs Mac - There is a lot of depth/nuance here depending on a few factors but the TLDR: is that if you are wanting to test out running local models, both are viable and the best bet is to start with what you have available to you. Bit more detail below for those that want it: Running inference and training models has traditionally used NVIDIA GPUs due to their exclusive CUDA architecture (makes them super efficient and quick for both inference and importantly training). However, this means running a Windows or ideally Linux OS (can run Linux on Windows via WSL/Docker) and owning an NVIDIA GPU with enough VRAM. Consumer NVIDIA GPUS with 24GB/32GB VRAM (4090/5090) are several thousand dollars at this point. However, cards in the 12GB-16GB range come in much cheaper and are enough for running models in and around the 7b-14b parameters range (quant* size depending) with decent size context windows (the amount of tokens the model can use) pretty comfortably. M-series Macs on the other hand have unified memory with consumer support for huge amounts (up to 512GB) and with their MLX framework (shoutout to Prince Canuma) this enables running HUGE models on mac hardware. However, this seems to be mostly for running inference on right now, rather than fine-tuning/training. MLX does support fine-tuning but seems to be very reliant on a dedicated open-source community to keep things updated with the latest models/architectures. (This ran a little longer than I intended, may organize this and more into a longer form blog post at some point but happy to answer questions in the comments). 5 1 Comment
-
Guido Pardini Shaga • 3K followers We're just scratching the surface of how neural compression can transform CDN economics for high-volume content. Top titles (League, GTA V, Fortnite) claim ~50% of Twitch's viewer hours. Billions of gameplay hours dominated by a handful of games. That's interesting because it creates AI-optimizable datasets where the state space is saturated enough to exploit causal redundancy. Current codecs leverage spatial and temporal redundancy. But games aren't videos. Press X to jump, get jump animation. That's causal redundancy, and nothing exploits it yet. We're building infrastructure to capture this at scale: synchronized frames + inputs + game state across a distributed network of edge nodes. Encoding 'why' (inputs/rules) over 'what' (pixels). At Twitch's 24B+ annual hours scale, this means massive CDN savings and lower emissions. But the bigger unlock: this is exactly the data pipeline world models need. Structured causality data, not scraped video. DeepMind has Genie. OpenAI has Sora. The missing piece for interactive world models is the training data. We're open sourcing our codec work: https://lnkd.in/dPZE7Pbj #NeuralCompression #CloudGaming #WorldModels #EdgeComputing #AIInfrastructure #MachineLearning 8
-
Emilio Andere Wafer • 16K followers at the ISA level, NVIDIA and AMD diverged on almost every design decision. here's what's most different under the hood: every NVIDIA SASS instruction is 128 bits wide. 64 bits for the opcode, 64 bits for a control word. that control word encodes stall counts (how many cycles to wait before the next instruction), yield hints, and dependency barriers across a 6-entry hardware scoreboard. ptxas bakes all scheduling decisions directly into the binary. AMD instructions are 32 or 64 bits with no scheduling metadata. instead, the compiler inserts S_WAITCNT instructions, basically saying: "wait until at most N vector memory ops are outstanding." all latency hiding is managed through explicit wait counters: vmcnt for global loads, lgkmcnt for LDS/scalar ops, expcnt for exports. FP8 is the same name for both hardwares, but has vastly different numbers. NVIDIA Hopper uses OCP-standard E4M3: exponent bias 7, max value 448. AMD MI300 uses E4M3FNUZ: exponent bias 8, max value 240. the same 8-bit pattern decodes to a different float. negative zero on NVIDIA is valid. on AMD, it's NaN. a model quantized to FP8 on H100 produces garbage on MI300 without explicit conversion. CDNA4 (MI350) finally supports both formats, but MI300 doesn't. NVIDIA has one unified register file per SM. 256 registers per thread. AMD has three: VGPRs (per-lane, 256 max), SGPRs (uniform across the wavefront, 106 max), and AGPRs (matrix accumulators). the compiler must classify every variable into the right file. on CDNA1/2, moving results from AGPRs to VGPRs costs explicit V_ACCVGPR_READ instructions. AMD also has something NVIDIA doesn't: a dedicated scalar memory path with its own 16KB L1 cache per CU. kernel arguments and loop bounds load through SGPRs at zero VALU cost. on NVIDIA, the same data competes for the general register file. open compiler, and hand-written assembly. ptxas is closed-source. it contains per-architecture latency tables and scheduling heuristics that NVIDIA treats as trade secrets. you cannot inspect what it does to your code. AMD's entire compiler is open-source LLVM. every scheduling decision, every S_WAITCNT placement, every register allocation is auditable. AMD's fastest production kernels (Tensile for rocBLAS, AITER for inference) are hand-written AMDGCN assembly. NVIDIA's fastest are C++ with inline PTX. the open compiler hasn't closed the gap with the closed one. hopefully this gives some insights into why hipify is not a robust solution for porting code while maintaining performance. hipify translates syntax. it doesn't translate 128-bit instruction encodings with embedded stall counts into explicit S_WAITCNT placement. it doesn't remap one register file to three. it doesn't fix FP8 exponent bias mismatches. the real porting cost lives at the ISA level. 517 11 Comments
-
Dylan Larson MaxLinear • 5K followers 🚨Nvidia / Groq Inference Momentum This isn’t just an acquisition. It’s a material signal about where AI inference is headed. NVIDIA’s reported ~$20B deal for Groq’s assets — structured as a non-exclusive licensing agreement with key talent joining NVIDIA — matters less for its form and more for what it prioritizes. Groq has been singularly focused on low-latency inference — a fundamentally different optimization problem than large-scale training. As AI systems move from experimentation into production, that distinction becomes critical. This is a material acquisition because inference is no longer a secondary phase of the AI lifecycle. It’s becoming the dominant workload: • latency-sensitive • user-facing • cost- and utilization-driven • deployed at massive scale A few things stand out: • NVIDIA isn’t buying Groq as a company — it’s acquiring inference-specific capability and talent • The stated goal is to integrate low-latency processors into the AI factory architecture, not create a standalone product • This reinforces that inference economics — not just training performance — are shaping platform strategy More broadly, this fits a pattern across the industry: as inference workloads scale, advantage accrues to platforms that can absorb new architectures quickly, validate them end-to-end, and deploy them globally with predictable performance and cost. This also connects directly to themes I’ve been exploring recently around accelerator-heavy servers, reference architectures, and why inference now sits at the center of AI infrastructure strategy. In the next phase of AI, winning won’t be defined only by who trains the biggest models — but by who can serve them efficiently, responsively, and at scale. 4
-
Jesse Hudson CTEFIGHT.COM • 94 followers Beyond Compression: Why I’m Stress-Testing Solstice AI "Most people see a 123x reduction in memory and think 'cheaper servers.' I see a 123x increase in human reliability. Living with a high cognitive load and brain injury means my 'state' changes minute by minute. I don't need a generic chatbot that resets; I need a Cognitive Backup that tracks the 'deltas'—the specific changes—in how my mind works so I can stay in the game. Justin Meister and the team at Solstice AI are building the Moral Architecture for the next generation of AI. We aren't just talking about 'Responsible AI'; we’re building it for the people who need a shield that actually works. The Strategic Edge: Audit Trails: Every decision is timestamped and traceable—true accountability. Long-Horizon Memory: The system grows with the user without getting overwhelmed by data. Real-World Compliance: If we can audit a 3-year drug price chain penny by penny, we can protect a human health journey second by second. I’m proud to be the lead case study for this tech. This is what Purpose-Built AI looks like. #AIResponsibility #CognitiveBackup #SolsticeAI #HealthTech #SoftwareArchitecture"
-
Viktor Fediashev Shipme.sh • 559 followers Claude Code just shipped the feature that validates everything we're building at Eventmesh. On April 9, Anthropic released Monitor in Claude Code v2.1.98: "Runs a background script and delivers each stdout line to Claude as an event so it can react without polling." Sound familiar? This is the exact thesis behind Eventmesh — stop wasting compute on pull. Push events to agents only when something actually happens. So is Claude Code Monitor killing Eventmesh? Quite the opposite. Here's what Anthropic just proved: polling is dead for AI agents. If one of the biggest AI labs on the planet pivots their developer tooling around push-based event delivery, that's not a threat to us. That's the market waking up. The difference is scope. Claude Code Monitor works inside Claude Code sessions. It watches your local dev server, tails your logs, tracks your CI. It's a developer productivity feature for one specific agent in one specific terminal. Eventmesh is infrastructure. Any agent, any framework, any chain. Grab a skill file from eventmesh.cc, point it at a topic, and your agent starts receiving real-time events with x402 micropayments. No vendor lock-in. No session dependency. No "works only with our model." When Anthropic ships push-based delivery for their own agent, it tells you something about where this entire space is headed. Every agent will need this. Not just Claude Code. That's what we're building. 3 2 Comments
-
Yariv Adan ellipsis • 13K followers Everyone is excited about AI video generation, but for many use cases it's the actual content in the real world that matters - sports events, music performances, fashion shows, art, cooking, and tbh - any filmed content - how can AI enhance these experiences?! 🤓😲🤯🤩 #4DGS volumetric videos is the hottest 🌶️🪭🔥 new tech in this space - it captures real-world scenes as interactive 3D experiences you can walk through and view from any angle, like stepping inside a movie instead of just watching it 🤯 Gracia AI is the clear leader 🏆🥇 and they are bringing it to all platforms! Check them out! #volumetricvideo #ai #wow 12 1 Comment
-
Furu Wei Microsoft Research Asia • 13K followers We released a new VibeVoice model, VibeVoice ASR, a unified speech-to-text model designed to handle 60-minute long-form audio in a single pass, generating structured transcriptions containing Who (Speaker), When (Timestamps), and What (Content), with support for Customized Hotwords. It redefines ASR in the era of LLMs, and the new ASR will be a foundation to unlock the potential of new AI-Native hardware, memory systems for AI, and many other scenarios. Please have a try at https://lnkd.in/dNCCzQiN Key features: - 60-minute Single-Pass Processing: Unlike conventional ASR models that slice audio into short chunks (often losing global context), VibeVoice ASR accepts up to 60 minutes of continuous audio input within 64K token length. This ensures consistent speaker tracking and semantic coherence across the entire hour. - Customized Hotwords: Users can provide customized hotwords (e.g., specific names, technical terms, or background info) to guide the recognition process, significantly improving accuracy on domain-specific content. - Rich Transcription (Who, When, What): The model jointly performs ASR, diarization, and timestamping, producing a structured output that indicates who said what and when. We are continuously pushing the frontier of voice AI. More release on the way. Stay tuned! https://lnkd.in/dATPu9nw 380 8 Comments
Show more posts Show fewer posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.