ramannanda.bsky.social @ramannanda.bsky.social · 05/10/2026Inflation is making people spend more for basics so spending is up? 000
ramannanda.bsky.social @ramannanda.bsky.social · 05/10/2026I am the house(casino) and I go bankrupt. 000
ramannanda.bsky.social @ramannanda.bsky.social · 02/10/2026Open models absolutely play into this. 020
ramannanda.bsky.social @ramannanda.bsky.social · 29/09/2026Astra has approx 192 experts with around 12 experts per token and sonnet has and is likely a dense model with 400b params. 041
ramannanda.bsky.social @ramannanda.bsky.social · 13/09/2026CSA2, etc. This is such a change from initial transformer architecture, all to efficiently absorb a lot of context and produce tokens efficiently and faster, while keeping kv memory super low. 011
ramannanda.bsky.social @ramannanda.bsky.social · 13/09/2026The deepseek guys are producing very compute frugal models, at one point one has to think, how can US labs efficiently compete. The causal encoder decoder with 8b/16b per token active and 890 bytes per token of kv cache in total across all layers,Engram(associative memory)192b params, SWA, 1/2 110
ramannanda.bsky.social @ramannanda.bsky.social · 27/08/2026Nvidia is the better option. Blackwell has so much advantage with TMA and pipelines. 010
ramannanda.bsky.social @ramannanda.bsky.social · 19/08/2026Well intervention shows gold getting a bid too. 000
ramannanda.bsky.social @ramannanda.bsky.social · 15/08/2026Indeed the 3.8 27b variant with fp8 is very good at tool calls, so for agentic use cases. It is a major step up. 000
ramannanda.bsky.social @ramannanda.bsky.social · 14/08/2026This strength show up in US treasury yields? 000
ramannanda.bsky.social @ramannanda.bsky.social · 11/08/2026Same both after covid and vaccination. 030
ramannanda.bsky.social @ramannanda.bsky.social · 31/07/2026They should focus on 30 year yield instead. 000
ramannanda.bsky.social @ramannanda.bsky.social · 31/07/2026Oh so that is what this escape means 😂 140
ramannanda.bsky.social @ramannanda.bsky.social · 29/07/2026So we keep rates steady lol see the 30 year tank. 020
ramannanda.bsky.social @ramannanda.bsky.social · 27/07/2026Could it have to do with China developing duv equipment? 101
ramannanda.bsky.social @ramannanda.bsky.social · 26/07/2026Why would it be both a leak and nonsense. 010
ramannanda.bsky.social @ramannanda.bsky.social · 26/07/2026Ends with just kidding. Jokes that need explanations are the best kind. 011
ramannanda.bsky.social @ramannanda.bsky.social · 26/07/2026Assuming people have morals, they wouldn't. 000
ramannanda.bsky.social @ramannanda.bsky.social · 13/07/2026Google created a new model for tabular data for this, yet to be released paper is out there though, research.google/blog/introdu...research.googleIntroducing TabFM: A zero-shot foundation model for tabular data 010
ramannanda.bsky.social @ramannanda.bsky.social · 11/07/2026Tempted :-), it is a nice version of minimalistic interior. 000
ramannanda.bsky.social @ramannanda.bsky.social · 11/07/2026Sonnet 4-6 is both faster, cheaper, and more effective for subagents; Sonnet 5 hits my harnesses' time-budget guards, never completing its output. #anthropic 000
ramannanda.bsky.social @ramannanda.bsky.social · 09/07/2026How many times you get to call the grok AI moron, is that metric reducing? 000
ramannanda.bsky.social @ramannanda.bsky.social · 30/06/2026Very expensive to run inference for trillion parameter models locally, inference demand is there for speed, but SLMs can run locally or you could quantize open weights models and run them on a significant local hardware. 110
ramannanda.bsky.social @ramannanda.bsky.social · 29/06/2026That 1b figure keep quoting doesn't include all the upcoming compute charges. 000
ramannanda.bsky.social @ramannanda.bsky.social · 25/06/2026Till how long is that gm of 84 percent sustainable? 000
ramannanda.bsky.social @ramannanda.bsky.social · 25/06/2026They are going to lower them eventually when this memory bubble bursts, 84 percent gross margin from 30 is not sustainable. 000
ramannanda.bsky.social @ramannanda.bsky.social · 12/06/2026Reuters reported that it was not Iran that attacked Indian ships. 101
ramannanda.bsky.social @ramannanda.bsky.social · 08/06/2026You want privacy nightmares? I know just the gadget for it. 010
