Sabitlenmiş Tweet
What does it look like to build AI that keeps power in the hands of the people? This is what @WorkshopLabs is about.
Starting today, we'll be sharing some of our work to this end.
English
Rudolf Laine
690 posts
Rudolf Laine
@LRudL_
What I'm doing: https://t.co/7tVMLt1OwN For my writing, see: https://t.co/GwKY6jk3Tl
Katılım Ağustos 2022
380 Takip Edilen2.4K Takipçiler
@TomDavidsonX But AI numbers will likely explode faster than human numbers?
Also direct copying of AIs means diffusion of cultural learning could happen much faster.
English
I think of cultural learning as a crucial part of the intelligence explosion, but don't think it could explode by itself
Historically, human cultural learning was super-exponential bc it was coupled with population growth and improvements in education. Without that, it would have fizzled out due to ideas getting harder to find
Similarly, for ai-driven cultural learning to be explosive it will need to be coupled with growth in the number of AI systems or their intelligence
Tyler John in SF 🇺🇸@tyler_m_john
I wrote an essay on the possibility of an intelligence explosion via cultural learning. Why are humans smart? Because we built a body of knowledge over 10,000 generations. This is the only form of intelligence explosion that's ever happened, so it could happen with AI too.
English
@tyler_m_john Underrated point! I wrote a series on accumulation of knowledge as a surprisingly core part of intelligence in humans and future AIs (with cultural learning a special case): rudolf.website/wisdom1/
English
I wrote an essay on the possibility of an intelligence explosion via cultural learning.
Why are humans smart? Because we built a body of knowledge over 10,000 generations. This is the only form of intelligence explosion that's ever happened, so it could happen with AI too.
English
A key q that splits people: is the key risk accident-like (we have some probability of stumbling into a catastrophe that would consume the world despite wide human resistance), or is it about the loss of power?
If accident-like: you just want to keep P(accident) low, and this often looks quite technocratic.
If loss of power: you need to think about incentives & political economy & who has leverage
@RichardMCNgo often makes the point that many tech people are blind to power concerns
English
@JacquesThibs @WorkshopLabs @LRudL_ Yes. AI safety is broad, but there seems to be more emphasis on loss of control and rogue actors than on concentration of power. We wouldn't want to democratize nuclear weapons, but that probably won't be a good analogy for AGI.
English
While I am for the automation of factories so that we can accelerate prosperity, there is a massive elephant in the room.
If people lose their jobs and economic power over the government, the government has less incentive to work for the people. This is a dangerous situation.
Bernie Sanders@BernieSanders
Jeff Bezos, worth $234 billion, plans to replace 600,000 Amazon workers with robots. Now, he wants to spend $100 billion to fully automate not just his warehouses, but factories in the U.S & other countries. Oligarchs are waging all out war against workers. FIGHT BACK.
English
Rudolf Laine retweetledi
@AddieF38654 @UnslothAI Third: how we made post-training and inference happen so privately that even we can’t your data.
Very excited about this! Hopefully it becomes an industry standard
x.com/WorkshopLabs/s…
Workshop Labs@WorkshopLabs
Letting a provider see all your data is the price of admission for AI. We're changing that. Introducing Silo, the first private post-training and inference stack for frontier models, with hardware-level guarantees that we can’t see your data. Privacy without compromises. 🧵
English
@AddieF38654 @UnslothAI Second: our own training stack for Kimi K2 Thinking, competitive on both price and performance. See below for how we got it to be so fast
x.com/WorkshopLabs/s…
Workshop Labs@WorkshopLabs
Introducing Trellis for Kimi K2 Thinking. It's post-training code that's 50x faster than the best single-node open-source version and 2x cheaper than training APIs. After safety testing, we're open-sourcing it, giving builders the best tools to customize a frontier model. 🧵
English
What does it look like to build AI that keeps power in the hands of the people? This is what @WorkshopLabs is about.
Starting today, we'll be sharing some of our work to this end.
English
Link to full post with all the details: workshoplabs.ai/blog/private-p…
English
Incredibly happy to have collaborated with @TinfoilAI. We’ve been a very demanding customer, but @t0nyav & co. have been incredibly helpful. Also shout-out to Daniel on our team who joined me in a mad sprint to extend privacy across our entire platform.
English
Giving up your data to AI providers isn’t necessary, but a choice. At @WorkshopLabs we chose differently and built Silo: post-training and inference for customized models, all without us being able to see your data.
How we did it: x.com/WorkshopLabs/s…
Below: why we did it
Workshop Labs@WorkshopLabs
Letting a provider see all your data is the price of admission for AI. We're changing that. Introducing Silo, the first private post-training and inference stack for frontier models, with hardware-level guarantees that we can’t see your data. Privacy without compromises. 🧵
English
@ryanyang0 @WorkshopLabs Tinker is not open source so we don’t know for sure, but also their shared-clock-cycle forward+backward across concurrent users architecture is quite different from us
English
@LRudL_ @WorkshopLabs i see and is gain bc you know it's kimi k2 and you can optimize for that specifically vs. or something else
English
Introducing Trellis for Kimi K2 Thinking.
It's post-training code that's 50x faster than the best single-node open-source version and 2x cheaper than training APIs.
After safety testing, we're open-sourcing it, giving builders the best tools to customize a frontier model.
🧵
English
@ryanyang0 @WorkshopLabs We don’t give you low level primitives like individual forward+backward, but instead you submit an entire run at once.
We do have some unique features we’re cooking up though, to be announced soon!
English
@WorkshopLabs the tinker comparison is crazy??? are they semantically equivalent?
English
@austinc3301 @mimi10v3 Decent, interestingly blames lots on losing the gold standard, might be in your water supply already if you read progress studies stuff and understand & agree w statements like “the Apollo program was a modern cathedral that serves a near-religious cultural purpose”
English
Tremendous alpha in iterating on data and evals for the 20th time, and resisting the siren call of being nerdsniped by fancy fun algorithms
English
The goals of all stakeholders can be aligned either through uniformity, or there being one stakeholder - you - and you having the power to build the system for yourself
Vibecoding does this for software - your software serves you. Next stop: AI
Julian@julianboolean_
"Software is by default thoughtless, and that allows it to be thoughtlessly cruel." recently i've come around to thinking that the most important factor in the functionality of a system is not technological capability or human competence, but simply alignment: whether the goals of all stakeholders are actually the same and whether they are properly incentivised towards the shared goal everything else is downstream of getting this right
English
@yudhister_ Credit to @AddieF38654 , who spearheaded the entire training stack project!
English
training is hard! glad @LRudL_ & co have pushed SOTA by leaps and bounds
Workshop Labs@WorkshopLabs
Introducing Trellis for Kimi K2 Thinking. It's post-training code that's 50x faster than the best single-node open-source version and 2x cheaper than training APIs. After safety testing, we're open-sourcing it, giving builders the best tools to customize a frontier model. 🧵
English