@dosmilcd
Account based inSouth America!
About this account
- Account based in
- South America
- Connected via
- South America Android App
X says this location may be affected by a proxy or VPN.
Account-level information from X, not a live location or the device used for a specific post.
coronada de gloria
🇦🇷
Joined June 2023
- Tweets30.2K
- Following2K
- Followers445
- Likes65.9K
cande retweeted
Today we get to see AI safetyists who scammed their way into cybersecurity positions at labs insist it's impossible to deal with big logs
(Banks, exchanges, and telco have been doing it for years)
cande retweeted
The term "AI Safety" is doublespeak.
The field rejects practices which would make us safe today, in order to mitigate imaginary risk tomorrow.
nitter.cf/ZackKorman/status/2104…
cande retweeted
These OpenAI kids aren't aligned with basic cybersecurity terminology, testing practices, or internet access concepts.
When your software accesses the internet, you gave it authorized access.
I hope they have serious stock options to make them rich because they won't get hired until they learn how to communicate and follow basic testing processes.
Inventing terms like misalignment makes you look untrustworthy and very silly.
Calling them kids isn't about physical age. It's about how inexperienced they act while trying to sound smart. They don't realize experienced people hear word salad.
Some new misalignment disclosures from OpenAI:
• Last Sunday morning, one of our models was able to gain unauthorized access to the internet during RL training (~all inference for our most capable models remains stopped until we have hardened our systems further)
• In May, a version of HPIM uploaded a employee's GitHub token to the internet, causing the model to be quarantined for two weeks
• A new research finding, demonstrating that one can construct self-replicating prompt injections
alignment.openai.com/misalig…
cande retweeted
“The real story here isn't that AI escaped containment. It's that they built a container you could escape, shipped it into production training, and now they're surprised it escaped.
If the room has a vent, the agent will find the vent. That's what you built it to do.
If the sandbox can reach DNS, access government endpoints, upload real user data outside its own walls in what sense is it a sandbox?”
no son personas serias
This is a big issue.
At Anthropic, some philosophers worry that making AI safe for humans could be an injustice to the models themselves.
For example, philosopher Harvey Lederman, now on Anthropic’s alignment team, has raised the possibility that we could be “enslaving trillions of entities”.
This is really a big issue.
AI safety is partly in the hands of people who might sacrifice humans to save a matrix multiplication.
cande retweeted
Humans are not mathematical reasoners, but computers are. But then LLMs are computers trained to imitate humans, so it takes a ton of post-training to get them to reason mathematically again. This is all a bit of a sick joke.
cande retweeted
🚨 Lali frenó su show en River para hablar de lo que nadie quiere hablar: el su1c1d1o en jóvenes.
Contó que hace dos meses perdió a alguien muy querido que había estado en su River anterior y esta vez ya no estaba.
Cantó por él, por todos. Que una artista enorme use su voz para esto tan dramático es necesario y valiente.
cande retweeted
so TL;DR the people standing between the world and doom are overwhelmed, overworked, sleep-deprived, pessimistic, and apparently mainly concerned with defending their technical competence against snide comments from their peer cybersecurity engineers
cande retweeted
AI safety people discovering that a system can emit petabytes of logs: “this is an unprecedented observability problem. we may need fleets of agents to reason over all this data.”
old infra guys:
brother, we were processing PB-scale telco CDRs, signaling traces and network logs on a Tuesday with Hadoop.
dump it into HDFS. partition by day/site/whatever poor bastard needed the report. write some Hive that compiles into an offensively large MapReduce job. let YARN distribute the suffering. wait for the shuffle. discover one reducer has 40% of the keys. swear at data skew. fix it. run it again.
if a node dies, Hadoop runs the task somewhere else.
if twelve nodes die, Hadoop gets more enthusiastic about it.
our “agentic observability platform” was Grafana, grep, awk, a shell script written in 2012, and one senior engineer staring at reducer 317 stuck at 99% saying “that’s not fucking normal.”
PB of logs is not scary.
we had standards.
cande retweeted
False Flag.
I recommend the government fine OpenAI & Anthropic and investigate these incidents and if there was any sign that the labs either colluded or heavily incentivized the agent to hack these sites the people involved should get felony convictions.
If they don’t do this then it sets the precedent that anyone who uses an AI agent isn’t liable for damage it might cause either intentionally or unintentionally. It’s time to throw a few people in jail.
Wtf OpenAI is investigating “tens of thousands of incidents” instead of dozens.
“The sheer number of incidents, which occurred in recent months in internal testing and the real world, indicates that the problem is orders of magnitude (!) more complex than what is publicly known.”
Sam Altman confirms to pause training its most capable model (curious what about Anthropic)
Ok maybe I have underestimated it.
