Greg Roodt

We’re excited to launch #MagicLayers globally at Canva 🚀🚀🚀 This is currently the most advanced image-to-layer decomposition model in the world-purpose-built for design-native content. ⚡ 20–200× faster than Qwen-Image-Layer 🎨 Super strong on designs canva.com/newsroom/news/…

@SemiAnalysis_ Source? Link?

NVIDIA MOAT ALERT: The performance of BLACKWELL increased 3.25x in the span of just 4 months.  At iso-interactivity of 95 tok/s/user, b200 deepseek fp4 perf went from 400tok/s/gpu to 1300tok/s/gpu

Feels like everyone making their own agent stumbles across the same primitives and thinks they solved something Let me save you some time (read this, it's funny and useful): - You're going to make an agent - You're going to run it on benchmarks > It's going to suck - You're going to make a tool to analyze traces - You're going to say this helped you > It wont work - You're going to think about role based agents for solving a single task - You're going to make a workflow for solving a benchmark > Both will work. Neither are generalized - You'll think you made it > It will be nearly unusable by an end user > Back to square one - You're going to realize you're stuck with a for loop - You're going to think about swarms > In swarms single agent usability doesn't matter - But wait you need a task manager - But wait you need a merge queue - But wait you need compression for long jobs > Compression is a foot gun - But wait now you need an agent to manage it all - But wait now you need something that checks to make sure the manager is managing - You're going to go back to single loop agents - Well, subagents seem like the way to do all of this > Bam! Plot twist: subagents are hard to do well - You're going to think "Hmm well subagents isolate context" because said so - You're going to start to look at other agent implementations > How have they all solved compaction, multi-agent, task management, memory etc.? - You're going to realize it's all just tradeoffs, but most of them have only one side people care about - "Oh it's all just context engineering" > Yep. But it has to be good and it has to be general. > Back to the starting loop. Rinse and repeat. Congrats. Keep it simple. Keep it general.

@jhleath Does archil support GPUDirect Storage?

about 10 years ago, when I was a fledgling engineer at Amazon EFS, we had a customer who’s entire data was stored as ONE BILLION kilobyte-scale json files in the root of their file system. Didn’t work well on EFS since then, it’s been my mission to make small files functional for users

For anyone trying to understand what Elon is talking about. Mass Driver en.wikipedia.org/wiki/Mass_driv…

I remain unconvinced that any companies need “realtime data”. The only place it makes sense is your observability stack, and you’re probably using datadog, honeycomb, or otel for that already. Everything else: use an OLTP database in product and a data warehouse for offline.

Forget MCP, what I want is an industry consortium backed standard for a JSON API protocol for talking to LLM providers Everyone building half-baked copies of the OpenAI Chat Completion API worked until it didn't any more!

everyone's constantly posting the meme about having a bunch of different agent rule files while the real nightmare continues to be totally ignored: - openai just dropped responses api that breaks every single existing agent architecture - anthropic format was the universal translator (superset of openai completions), now it's obsolete - every provider has different message shapes, tool calling patterns, reasoning hydration - cline has anthropic baked into disk storage, 30+ providers, core interfaces - migration would total architectural hell who gives a fuck about .cursorrules vs agents md when your reasoning traces disappear between api calls and your entire codebase assumes one message format that's no longer the superset? can we please standardize on a future proof llm API standard?

Anyone else waiting for Zuck to announce that Meta has hired Soham Parekh?

Superhuman is being acquired by @Grammarly! 💜💚 Together, we will build the AI-native productivity suite of choice 🥇 We will invest even more deeply in AI and email, reimagine chat and collaboration, and build AI agents that unlock a whole new way of working. More below 👇

@mitsuhiko I've also encountered groups who rampage around a company trying to replace all of the the custom, wonky shaped homegrown wheels with "don't reinvent the wheel" external products and libraries... without taking the time to understand why the wonky wheels are that particular shape

@pymhq Congrats to everyone one theahnch! Is it too early to ask about performance expectations on Bedrock vs 1P?

Teams are currently having happy hour with the Anthropic offsite crew. Cheers #Claude4 🥂

@pymhq Trainium?

Check out my current team’s latest delivery: ~2x model inference acceleration.

New Evals API I’m excited to share a new API for logging evals with W&B Weave. EvaluationLogger - log_prediction - log_score - log_summary Our design goal for this API was to get out of your way and build the most flexible eval API out there, inspired by wandb.log, which our @weights_biases users love. - No hidden logic, you control the eval loop and what you log - Easy to integrate into existing evals with any model or framework - Log and version everything so you don’t accidentally compare incomparable things - Easy to query - Works with our existing comparison UIs We'd love you to try it out and share your thoughts on it. link below

@OfficialLoganK Are you going to simplify the cache API? It's the most awkward one to use at the moment compared to competitors.

Context caching updates in the Gemini API: - ✅ Added support for 2.0 Flash - ✅ Added support for 2.5 Pro Preview - 📉 Reduced min context size from 32K down to 4K Much more to come still, please send any feedback on the experience!

@OfficialLoganK Congrats on the release! Gemini is crushing it across the board recently. Do you know if there are any plans to improve the caching semantics? It's more awkward to use than your competitors.

The Gemini 2.5 series is the world’s best model lineup 👍

Agents are just selecting a DAG of tool calls aren't they? So we're doing query planning, optimization and execution again aren't we? It's SQL isnt it? The answer is SQL. It's always SQL. AQL: Agent Query Language.

Remember xgboost?

At @Tailscale we just raised a $160M Series C, marking our official transition to a growth stage company. Hold on to your socks. tailscale.com/blog/series-c

@pymhq Bedrock is so good! Would you estimate the ETA is days? Weeks? Months?

Llama4 is single node, and Bedrock’s onboard coming soon.