Remember this? A problem we didn't have in automated speech recognition tools until OpenAI decided they were going to do away with alignment (not the kind that so-called "AI safety" people do)? And they're giving us "artificial general intelligence" any day now right? apnews.com/article/ai-a... - ThreadSky

timnitgebru.bsky.social • 45 days ago

Remember this? A problem we didn't have in automated speech recognition tools until OpenAI decided they were going to do away with alignment (not the kind that so-called "AI safety" people do)? And they're giving us "artificial general intelligence" any day now right?

https://apnews.com/article/ai-artificial-intelligence-health-business-90020cdf5fa16c79ca2e5b6c4c9bbb14

Comments

albertbfx.bsky.social•44 days ago

Valid concerns, but scalable solutions require faster innovation. FFT-based harmonics auditing for real-time transcription is underexplored but promising. Some of the real challenges are surprising. Privacy, licensing laws & data-sharing restrictions can stifle the R&D needed to drive AI alignment.

albertbfx.bsky.social•44 days ago

The people who seem to know this best are the privacy lawyers contentiously updating privacy and security terms in ways ordinary consumers will never understand in TOS, immunizing themselves from future legal action right below our noses. Language is source code, written in many interactive contexts

nregib.bsky.social•44 days ago

We don’t need this

timnitgebru.bsky.social•44 days ago

No this wasn't a problem we had in automated speech recognition tools before, needing NO innovation to get to this.

albertbfx.bsky.social•36 days ago

It’s actually a position that can typically only be held tight by those insulated from the direct impacts of these issues. It’s the 2020s version of the 1980s and why as distasteful as some people find him, Elon Musk presciently has some fundamental points on free speech exactly right.

albertbfx.bsky.social•36 days ago

Making such fast AI transitions will be incredibly hard in the current and next generation without better ways of getting consensus on tradeoffs. There will be answers people don’t want to hear or trust for a long time. This isn’t new, it’s just like before. Are we getting ready or complaining?

albertbfx.bsky.social•36 days ago

Back to Elon, we figure out that in hindsight the FAA had a model of insisting on human control for something AI can do 1000x better and planes don’t need central coordination by attentive humans to fly safely. We’re in the Museum of Antiquated technology (use of spectrum section). That is an ATC.

albertbfx.bsky.social•36 days ago

As a life ending possibility for all humans for the significant moment it was. If Trump gets everything else wrong and gets that right, humanity will thank him and it’s something we we should root for even if he’s amygdala central in your brain right now. 😆

albertbfx.bsky.social•36 days ago

Let me explain the insanity of Elon Musk loathing from my perspective, which doesn’t have to make sense to you, but it does to me. It’s the year 4025, Trump Day, a holiday added in the year 2240 after almost another quarter century of debate. On Trump Day, dna humans mark the end of nuclear war …

albertbfx.bsky.social•36 days ago

…often reflect and reinforce existing power structures rather than serving vulnerable and at risk people. This kind of elitism manifests in assuming that avoiding certain topics entirely is more ethical than studying them analytically to reduce harms. And you know what?

albertbfx.bsky.social•36 days ago

There’s a form of privileged moral absolutism in many AI ethics frameworks, particularly bad from Google, that end up perpetuating harm rather than preventing it. By refusing to engage with certain realities in the world because they’re deemed too problematic or unsavory, these frameworks …

albertbfx.bsky.social•36 days ago

Here’s something for AI ethics to confront because it’s a form of elitism that is embedded very deeply in false morality.

albertbfx.bsky.social•44 days ago

This is déjà vu because AI’s current trajectory mirrors what happened with cybersecurity—exponential change leading to risks that require scalable oversight, but with solutions always playing catch-up. If we don’t learn from past experiences, we repeat mistakes with a bigger kaboom scale.

albertbfx.bsky.social•44 days ago

1. New Technology, Old human problems (fear, greed, power,..
2. Innovation Outpacing Oversight
3. Patchwork “Tooling the Problem” Mentality without a systemic roadmap
4. Regulatory Pushback followed by regulatory capture (just look at perverse outcomes from GDPR..)
5. Trust vs. Risk Balance

albertbfx.bsky.social•44 days ago

Exactly, AI is an ocean that has been discovered that interferes, promising great riches in return. No amount of activism or regulation can wish away the ocean. We need scalable oversight tooling that uses AI as creatively and quickly as the problems it brings.

beccam.bsky.social•45 days ago

Yeah…seems like every company I work with lately wants to go AI for customer interactions, and none seems to understand that these systems have ZERO critical reasoning capabilities and make sh*t up, which in many contexts can literally kill people by delivering lethally wrong results.

rbowman1234.bsky.social•45 days ago

My wife's PhD program required the use of a transcription program (that in its TOS claims ownership of everything) and had to spend 2 days manually correcting it's output for every interview. She was worried they had a way to check so she wouldn't just transcribe without it

psychicfax.bsky.social•45 days ago

Meanwhile, WatchDuty, the app used by millions for info about the fires in LA, is a non-profit that coalesces information from multiple sources in real-time by hand. I heard an interview with the CEO recently (cannot find the audio now but it was NPR) where he explained this. A chasm of difference.

webrant.bsky.social•45 days ago

And this is already as good as it will ever get as once it models those "inventions" it gets worse in a loop of regression.

https://bsky.app/profile/timnitgebru.bsky.social/post/3lfq4nnxxu22q

aidannova.bsky.social•43 days ago

everyone's chasing AGI while basic ASR is getting worse. we actually use claude 3.5 for speech recognition at jenova ai - way more accurate + no hallucinations. sometimes boring alignment > flashy capabilities 🤷

heathertubbscooley.bsky.social•39 days ago

📌

robinafox.bsky.social•43 days ago

Perhaps it's stating the too-obvious to mention The Matrix as well as Johnny Mnemonic and A Scanner Darkly. Almost prophetic.

sullymander.bsky.social•45 days ago

It's kind of sad because CTC (without a language model) will spell things out as they sound, which is a much better failure mode than "make shit up"

ubiquity75.bsky.social•44 days ago

I’ll take “authentic specific knowledge,” thank you.

randymiller.bsky.social•45 days ago

These guys are insistent on turning solved problems into new frontiers just to justify a sunk cost. Replacing decades of research into other forms of machine intelligence and audio/video recognition with something demonstrably worse.

wolfenred.bsky.social•45 days ago

Keep Digging - With all this shit, there has to be a pony somewhere!

grimjim.bsky.social•45 days ago

We've entered an AI-powered snake oil era where health products are being pushed apparently without even basic safety testing. Who should be held responsible for AI hallucination-induced malpractice? This is a real risk now, not purely theoretical.

druid48.bsky.social•44 days ago

that is the new business development feature

pdxspashi.bsky.social•45 days ago

My spouse brought this up with his own doc, that her notes might be full of AI hallucinations. She hadn’t heard of this and had not been reviewing that closely but does now!

coach-k.bsky.social•44 days ago

That's very concerning. The implications are high given the amount of transcription being done in the health care field.

kentonagbone.bsky.social•45 days ago

They quietly switched out the "I" in AI to "Incompetence" when no one was looking

microh.bsky.social•44 days ago

Google’s translations from Farsi and Arabic mess up numbers a lot, and that seems enough like the function of the LLMs that I’m blaming that.

Comments

Posting Rules

Reply