GPT-5 might be farther off than we thought, but OpenAI wants to make sure it is safe.
See full article...
See full article...
OpenAI training its next major AI model, forms new safety committee
- Thread starter JournalBot
- Start date
These openAI posts always read like PR Puff pieces and not unbiased reporting. How is the above statement true outside of what OpenAI states?
Upvote
235
(244
/
-9)
All marketing noise until we can exactly how and what it is trained on.
Oh wait , "openAI" is not open. But then again, it's not "AI" either ;-)
Oh wait , "openAI" is not open. But then again, it's not "AI" either ;-)
Upvote
122
(143
/
-21)
This sounds more like an industrial strength CYA filter designed to keep Altman safe. If it was a part of the product or engineering orgs I'd be hopeful, but this seems more like a way to prevent information flowing from product and engineering to the board.
Upvote
144
(146
/
-2)
The rapid progress in LLMs, coupled with the equally stagnant progress in their actual ability to evaluate evidence, produce factual results, handle novel situations, and generally not lie ("hallucinate") with confidence shows that we're probably a century or more from AGI, and that AGI will not be primarily LLM-based.
How often do we still see "Sorry about that, let me try again:" followed by some different mistake, even for very basic questions?
How often do we still see "Sorry about that, let me try again:" followed by some different mistake, even for very basic questions?
Upvote
72
(108
/
-36)
Zero confidence in any form of self regulation, especially because effective and balanced regulatory schemes governing AI are a solved problem in jurisdictions like the EU, which have adopted a sophisticated and broadly applicable AI law.
Upvote
89
(93
/
-4)
Re. the delayed release of ChatGPT-5, it doesn't surprise me that LLMs will soon reach a limit of capability. You can only throw so much data at what is a relatively simple algorithm before you get diminishing returns - just like deep neural networks a few years ago.
If we examine the workings of the biological brain (the only extant AGI machinery we know of) we can clearly see a complex network of specific structures and meta-effects from neurotransmitters that, to my knowledge, don't have any equivalent in contemporary AI software. I think it's reasonable to expect that evolution's solution is well optimised despite the complexity, so why should we expect AGI in software to be any simpler?
I am a little surprised that multi-modal models haven't broken through the limitations of LLMs, they feel closer to brain structures to me. Perhaps plumbing outputs from different models together (forgive the metaphor) is proving more difficult than it sounds.
Or maybe there's nothing wrong with LLMs in principle and the problem is simply pollution of training data (see Dead Internet Theory).
If we examine the workings of the biological brain (the only extant AGI machinery we know of) we can clearly see a complex network of specific structures and meta-effects from neurotransmitters that, to my knowledge, don't have any equivalent in contemporary AI software. I think it's reasonable to expect that evolution's solution is well optimised despite the complexity, so why should we expect AGI in software to be any simpler?
I am a little surprised that multi-modal models haven't broken through the limitations of LLMs, they feel closer to brain structures to me. Perhaps plumbing outputs from different models together (forgive the metaphor) is proving more difficult than it sounds.
Or maybe there's nothing wrong with LLMs in principle and the problem is simply pollution of training data (see Dead Internet Theory).
Upvote
88
(103
/
-15)
Baking in conflicts of interest by having your release-it-now CEO on the wait-til-its-safe safety committee sure is a choice...
I guess giving their committee inherently misaligned incentives and power structures will keep their minds focused on them.
I guess giving their committee inherently misaligned incentives and power structures will keep their minds focused on them.
Upvote
109
(110
/
-1)
Me, too, although my suspicion is just that we don't have any good grasp on what modalities we actually need, and are rather just using what we have, coupled with LLMs' poor ability to characterize correctly. If they don't "understand" that the tokens they're "looking at" at the moment are essentially a "math" problem--or how to correctly rephrase it in its mathematical form--all the access in the world to things like Wolfram Alpha don't help.
Upvote
24
(28
/
-4)
How will the so called safety committee be able to act with any form of independence or object analysis of the potential risk associated with OpenAI’s products when the company’s super capitalist tech bro CEO is on the committee?
-edit- clarification
-edit- clarification
Upvote
88
(90
/
-2)
I feel like the "Safety and Security Commission" only exists to avoid spooking investors and keep the money flowing in. I doubt it will engage in what many of us would consider safety in regards to LLMs.
Upvote
60
(61
/
-1)
This whole thing reminds me of a proto-story that's been brewing in my mind for a while. It's about a convergence of technologies: over time, mankind shrinks processing as far as it can go at a mechanical level, and switches to chemichal processors. AI scientists want to take advantage of this, but feel it's too dangerous. At the same time, new research into wormholes discovers what scientists think is the ability to create a wormhole to an alternate reality where the world is undeveloped. The AI scientists seize this as an opportunity to claim the "new world" as their biochemical AI testing ground in the race to produce the first AGI.
Unfortunately, it turns out that the wormhole tech actually punches a hole through space and time, and the AI tech runs amok on the new world, destroying the planet... which turns out to be Mars a few billion years ago. And some of that tech hitches a ride somewhere else on an impact meteor....
Unfortunately, it turns out that the wormhole tech actually punches a hole through space and time, and the AI tech runs amok on the new world, destroying the planet... which turns out to be Mars a few billion years ago. And some of that tech hitches a ride somewhere else on an impact meteor....
Upvote
30
(34
/
-4)
The whole thing smells like bullshit. Not just the commitment to “safety,” but the claim that it’s somehow of critical concern.
LLMs can’t even output consistently valid information and fail at anything approaching skilled tasks. So it’s mighty convenient that the companies selling them talk about how their greatest fear is an LLM creating Skynet and causing a nuclear holocaust.
LLMs can’t even output consistently valid information and fail at anything approaching skilled tasks. So it’s mighty convenient that the companies selling them talk about how their greatest fear is an LLM creating Skynet and causing a nuclear holocaust.
Upvote
62
(70
/
-8)
Altman recently said that the capabilities of the GPT-4 model are an embarrassment. It was quite a big hint at the capabilities of an incoming model, so I’m really surprised that they have now just started training it.
Perhaps it means that they have started training a GPT-5 that will adhere to the new safety standards, but will still release a GPT-4.5 this summer following their old approach to safety?
Perhaps it means that they have started training a GPT-5 that will adhere to the new safety standards, but will still release a GPT-4.5 this summer following their old approach to safety?
Upvote
-17
(7
/
-24)
"saftey committee"? Is that like google saying they value our privacy?
Or facebook's committee - the one with ZERO power that gets overruled by the rest of the company 100% of the time? Just cause there is a "committee" doesn't mean it gets listened to or empowered.
Or facebook's committee - the one with ZERO power that gets overruled by the rest of the company 100% of the time? Just cause there is a "committee" doesn't mean it gets listened to or empowered.
Upvote
65
(66
/
-1)
Eventually I think we'll find that the ever increasing addition of guardrails eventually ends up with a straightjacketed LLM. I wonder, if in a few iterations, we end up creating a school of sorts to not only provide training data but the expected ethics around the use of that data, and what that may look like.
Not forcing certain behaviours, necessarily like we do now, but training them into the model?
Not forcing certain behaviours, necessarily like we do now, but training them into the model?
Upvote
3
(7
/
-4)
I'd take a guess they had to wait for new hardware to be procured, installed and validated.
Upvote
4
(5
/
-1)
Longmile149
Ars Scholae Palatinae
I will be absolutely shocked if this actually happens in a format that isn’t “OMG YOU GUYYYS OUR AI IS LIKE WAY TOO SUPER PWRFUL! WE SHOULD BE TOTALLY CAREFUL oh btw you can subscribe to new GPT4oTurbo (ASMR JOI Edition) for $27/month”
Upvote
49
(52
/
-3)
Safe for what? Fudging English papers badly? If they're close to a general purpose AI, it's nowhere evident from the current iterations. This looks like a backhanded humblebrag unconnected to reality.
Upvote
29
(35
/
-6)
Ah, that explains a lot of things. The discontinuation of the Zilog Z-80s, the expedition to asteroid Bennu, the delay in the new models, the endless 20-years-out-of-fusion-power, the sudden new interest in lunar manufacturing....in thirty years we're going to wake up to a new moon on the sky, the AI Overwatch powered by clean energy, every neuron a Z80 created by moon-based facilities from space-borne materials!
Upvote
15
(17
/
-2)
How's about they appoint a "don't break the law" committee which could keep them safe from I dunno legal exposure from torrenting a bunch of movies with subtitles to train a text to speech engine.
I'm calling it, their non text based stuff will get them nailed to the cross. Warner Bros is not NYT. You can't simply download "Her" from Warner Bros website with no DRM, for search engine indexing.
I'm calling it, their non text based stuff will get them nailed to the cross. Warner Bros is not NYT. You can't simply download "Her" from Warner Bros website with no DRM, for search engine indexing.
Upvote
25
(29
/
-4)
Every regulated industry has a disclosure element: ingredients, MSDS, financial records and corporate forecasts, results from scientific studies, etc... These all are intended to provide enough data to say: here's what we starting with, and what we are doing with that.
The closed nature of OpenAI's training data means we cannot evaluate where they are starting from--whether it's safe/fair/valid. Without that info, this committee is just a shield/show that's trying to convince the public to 'trust us' without evidence.
The closed nature of OpenAI's training data means we cannot evaluate where they are starting from--whether it's safe/fair/valid. Without that info, this committee is just a shield/show that's trying to convince the public to 'trust us' without evidence.
Upvote
42
(44
/
-2)
Count me among them, because it seems very obvious to me that a language model is not AGI, categorically and obviously, and that LLMs cannot be bootstrapped up to AGI and will not naturally lead to AGI or anything even credibly mistakable for same.
This is why LLMs piss me off. They're a novelty with some limited usefulness and a number of deep foundational flaws, but there's just enough meat there for it to feel genuinely revolutionary, thus making it feel more solid than bullshit like crypto and bitcoin and therefore perfect fodder for stock-pumping grift.
Upvote
72
(85
/
-13)
Yeah these things really have a “thoughts and prayers” vibe to them.
Upvote
22
(26
/
-4)
I think they do train new models constantly. The 4-turbo an 4o were both new smaller models that were meant to keep the capabilities of a much larger original GPT-4.
I actually suspect that 4o is not that much bigger than GPT-3.5-turbo based on the fact that they are making it free to all users.
Upvote
12
(12
/
0)
Agree that this new committee gives off heavy CYA vibes. Bear in mind that there is never any guarantee of good faith— will they ever actually do anything other than emit hot air?
Upvote
-1
(0
/
-1)
OpenAI is starting to take their naming cues from Capcom here:
ChatGPT 3.5
ChatCPT 3.5 Turbo
ChaptGPT 4
ChatGPT 4o
ChatGPT 4o Turbo
ChatGPT 3.5 Championship Edition
ChatGPT 3.5 Turbo Hyper Edition
ChatGPT 4o Alpha...
Upvote
34
(36
/
-2)
Of course Altman is on the “new” “”safety board””. How else is Altman gonna reach Elon’s status, if he cant grift with impunity!
Ah, truly wonderful times we are living in*
*Im gonna go dig a big hole in my backyard now, got the concrete coming in next week.
Ah, truly wonderful times we are living in*
*Im gonna go dig a big hole in my backyard now, got the concrete coming in next week.
Upvote
19
(25
/
-6)
Unless this committee has the ability to shut down the work being done, no questions asked, then the committee is worthless.
Upvote
14
(20
/
-6)
Longmile149
Ars Scholae Palatinae
I am personally still processing their incredibly revealing decision to say “what if Chat GPT…but fuckable?”
The obviously very proud of themselves showcasing of the flirty lady voice that just can’t get enough of the user is so quintessentially tech bro that I just can’t parse it all the way out yet. My mind refuses to fully engage with how deeply, depressingly stupid the entire company is, and, by extension, the whole VC tech set.
ETA:
I mean, for fuck sakes. Sam Altman made people go on live TV and introduce his fucking LLM waifu pillow to everyone like it is just completely normal shit and not a sad statement about the obscene epidemic of isolation and loneliness that he and his Silicon Valley buddies have worked so diligently to perpetuate.
Upvote
31
(43
/
-12)
telling the UN my nuclear arms business doesn't need regulation because I'm just 'doing physics for money'
Upvote
46
(47
/
-1)