Smit B Patel

@claude has real issues. You do not exhaust the limits and it still says you can't use it anymore until the limit resets.

@claudeai is forever down!

Giveaway time! Win a Claude Max 20x subscription for one month. To enter > follow @BleapApp > RT and like this post Winner will be selected in 48 hours. (fyi, you get 20% cashback on your claude, chatgpt and gemini subscriptions when using a Bleap card) Download the app via the link in our bio today and activate your virtual card in minutes.

@NandoDF Thanks for the share @NandoDF! Also, wonderful to know the scope to be a part of @MicrosoftAI, Cheers!

playground.microsoft.ai is now available in the USA and soon all over the world. Today we unveil our image generation models. Soon everyone will have access to many more models to play with in our playground. Enjoy! Wanna help build these models? JoinAITeam@microsoft.com - In particular, I'm looking for exceptional data, research and infra engineers.

@djhelbert @awscloud @credly Congratulations!

View my verified achievement from @awscloud. credly.com/badges/c72d279… via @credly

What I like most is the systems idea: keep expensive hardware busy by doing the right work in parallel, and use prediction plus caching to avoid idle time (Don’t let the no-work time go unused!) #AI #LLM #Inference #Systems #GPU #PerformanceEngineering

Reported results (on Llama-3.1-70B with H100 GPUs) show up to about 2x faster than optimized speculative decoding and up to about 5x faster than standard autoregressive decoding.

I read a new paper called “Speculative Speculative Decoding” (arXiv:2603.03251). The problem is simple: large language mathematical systems generate text one token at a time. That sequential loop becomes the bottleneck during inference, even on powerful GPUs.

@avnermay @tanishqkumar07 @tridao

Paper link: arxiv.org/pdf/2603.03251

What I like most is the systems idea: keep expensive hardware busy by doing the right work in parallel, and use prediction plus caching to avoid idle time (Don’t let the no-work time go unused!) #AI #LLM #Inference #Systems #GPU #PerformanceEngineering

Reported results (on Llama-3.1-70B with H100 GPUs) show up to about 2x faster than optimized speculative decoding and up to about 5x faster than standard autoregressive decoding.

They call this framework Speculative Speculative Decoding (SSD), and the optimized version is called Saguaro.

@VazeKshitij @deepigoyal Congratulations!

@deepigoyal You just made my weekend man

Temple has raised its first round. Friends and family. $54m. Post-money valuation of ~$190m. Every investor in this round is a founder friend or early-stage Zomato investor who wanted in, whether or not Temple ever makes it to market. But here's what gives me goosebumps – more than 30 Temple employees participated in the round, at par valuation. No discount. Their own money. That's the kind of belief you can't buy. We are assembling a dream team to build the ultimate wearable for elite performance athletes. Want in? Look up my last post.

@TSemenski Great job. Thanks @TSemenski

That's matrix multiplication in ML. Given these inputs and these weights, what are the outputs? Once you see it this way, you'll never forget it.

Most ML engineers use matrix multiplication every single day. Only a few of them can explain what it's actually doing. Here's how matrix multiplication REALLY works, explained visually so it finally clicks 🧵👇