Pinned Tweet
Non ascoltate me, ascoltate @repubblica - forse sarebbe ora che nPlan si attivasse anche in Italia per prevenire altri disastri? Per non parlare del ponte sullo stretto...
repubblica.it/tecnologia/202…
Imagine waking up one day and discovering you've invested in the cheems copycat of @nPlanHQ
A tutti i decerebrati che pensano che stia facendo propaganda a Repubblica, sto postando LA MIA AZIENDA. Lo so cosa fa e come funziona, e Repubblica ha scritto un articolo su di noi. Fatevi furbi.
Been at @MCO for 20 mins, only person in the @British_Airways checkin line. Just been staring at checkin attendants scratch their neck for all that time.
My DaTa is MOre VAluaBle To yoU tHaN YOUr pROducT Is tO Me
Stop. You sound like an idiot and you have just revealed to everyone you don't understand #ai as it is since after 2023.
This is the #ai benchmark we were all waiting for
Alan Mosca retweeted
It's wild to watch a $600B market swing in the wrong direction on news that each unit of compute is actually *more* valuable than previously appreciated.
DeepSeek’s cost-saving breakthroughs aren’t new ideas (MOE, MLA, fp8 training), but they’re the first to make them work together effectively.
* MOE (Mixture of Experts) increases training efficiency by activating only 37B of 600B model parameters, cutting compute by 80%.
* MLA (Multihead Latent Attention) speeds up inference by projecting attention into a smaller space and caching it.
* fp8 (8-bit floating-point) reduces memory use by up to 75% compared to fp32 and speeds up computation dramatically, allowing training on 2,048 H800 GPUs.
These concepts weren’t widely adopted together before because they weren’t seen as practical. This isn't some crazy inflection moment.
Calm down everyone. Did we learn nothing from the superconductor hype?
Markets overreact to the latest news, forget the past, and jump to conclusions.
Reality check on #DeepSeek:
* It’s not the 'end' of other #AI companies. Especially not @OpenAI
* One benchmark ≠ universal success.
* HuggingFace hasn’t even fully reproduced it yet.
* Next week, it’ll be another model.
* Progress in AI = more usage
Let’s stay grounded. Dip buying opportunity?
Most of #GovernmentEfficiency can be achieved by preventing delays and blowouts on megaprojects.
Smarter risk management and forecasting will do that.
@elonmusk GAO and DCMA are best placed to achieve this with the right tools!
The world doesn't need another hot take on #DeepSeekR1 .
Just shut up and get back to work.