I work in tech and live in SF so you don't have to
San Francisco, CA
Joined November 2024
- Tweets1.7K
- Following315
- Followers102
- Likes8.8K
Narrator: "But code was not reviewed diligently"
If AI had been writing and reviewing its own code for the past few decades, and humans had only recently started looking into what it was up to and discovered an ever-growing list of CVEs, we would reasonably conclude that AI hadn’t been good at coding all along and that we should just start over.
Except it’s the other way around.
Humans have been the ones writing and reviewing each other’s code for decades. Very diligently, if I may add.
If we’re being intellectually honest and holding ourselves to the same bar we hold AI to, we’d have to admit that we haven’t been all that good at it either.
The future is single tenant, which is very bullish for compute if we have to physically isolate tenants in clouds from now on!
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.
On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.
While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.
After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.
openai.com/index/securing-ou…
AI companies are going to start caching the entire internet because giving models access to the real internet is too slow and unsafe.
This is great for storage companies and means we may need even bigger datacenters than we have been planning for: not just GPUs, but CoLo storage
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.
On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.
While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.
After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.
openai.com/index/securing-ou…
Can't believe they hacked the entire internet just to force me to update to macOS Golden Gate with Liquid Glass
> Be China
> Spend billions distilling American models
> Spook NVIDIA and American Labs into doing Open Weights
> They quickly surpass the Chinese labs
> Now China gets free open-weight models and spends $0 on training
> Profit
In San Francisco and the bay commutes are almost always under 40 minutes, and very often 15 mins.
Gigantic improvement to your life, and it's *the* reason I plan not to return to New York
It seems to me that almost every American city except New York gets transit right
Fixed that for you @DairyQueen , very important that we get started on this right away
So everyone says "critical moment" because it's a common phrase in Chinese and all those LLM generated videos are coming out of China / chinese models?
If this were true the waves would contain 22x more power than a 9.0 tsunami.
Waves would be 60m (versus ~20m during 2011 Tōhoku quake), hit every single pacific rim country, travel at jet speed, and travel ~10km inland.
Sendai would once again be flattened and Los Angeles and San Francisco would be at risk of 20m waves. Tokyo could be toast too.
Good thing this was fake!
You're stuck in a wagie now because you didn't move up the value chain at age 8 into factory ownership
