But neural networks... they're an exercise in empiricism. We only ever understood that it works, never why. It's sort of a miracle that it doesn't produce buggy output all the time.
What do you tell people when they want to know why the miracles have stopped? Root cause: the gods are angry.
Well we wouldn’t have to guess if their PM wasn’t so pointlessly vague would we?
As someone who learned to read in part with the Commodore 64 user manual telling me about PEEK and POKE while I was actually in primary school, I think you're greatly overstating what primary school children write about in their holidays.
Snark aside, is their message vague? Sure. But more words wouldn't actually tell us more unless they also published a lot more about their setup — "we were casting float16 to uint16 on line 249 of server_config.c" isn't going to really help either.
Also, here's a recent security update from Apple to compare against: https://support.apple.com/en-gb/HT214056
I think your counter-example swings too far in the other direction. Nobody expects a git-diff of the fix, but a solid explanation of the whys and wherefore’s isn’t unreasonable. Cloudflare does, fly.io does, etc etc.
E.g.: Have a gauntlet of 20 moderate complexity questions with machine checkable characteristics in the answer. A couple may fail incidentally now and then but if more than N/20 fail you know something's probably gone wrong.
I used to be an SRE at Google and I wrote up internal postmortems there. To me, this explanation feels a lot like they are trying to avoid naming any of their technical partners, but the most likely explanation for what happened is that Microsoft installed some new GPU racks without necessarily informing OpenAI or possibly only informing part of their ops team, and that this new hardware differed in some subtle way from the existing hardware. Quite possibly that means a driver bug, or some sort of hardware incompatibility that required a workaround. Certainly, they would not want to be seen publicly attacking Nvidia or Microsoft given the importance of these two partners, so keeping it high level would certainly be for the best. Virtually. None of openai's customers would be able to use any further technical detail anyway, and they may still be working out a testing strategy that would allow them to detect changes in the hardware mix that unexpectedly cause regressions without necessarily any software deployments being involved.
For companies this size, with these valuations, everything the public is meant to see is heavily curated to accommodate all kinds of non-technical interests.
Sure, you can "step through" a neural network but all you'll ever see are arrays of meaningless floats moving around.
OpenAI moves so quickly with their product and doesn't seem to be shy about quickly pushing changes to prod. There are too many times to count that, using ChatGPT, I've seen the frontend changing before my eyes or seen responses change as backend modifications are pushed.
On one hand it's refreshing to see their lack of aversion to push changes. On the other hand, it leads to probably one of the most unpredictable experiences I've had using any consumer product.
I’m still using the original gpt-4-3.14 model.
Seems pretty clear. A good interpretation is that they had a test escape for certain GPU configs.
> In this case, the bug was in the step where the model chooses these numbers. Akin to being lost in translation, the model chose slightly wrong numbers, which produced word sequences that made no sense.
I liked this because when I first saw the example word salads I was so impressed by them - they look to be syntactically correct, but semantically they're gibberish. But knowing the basics of how LLMs choose the next token let me imagine some bugs where the "lookup table" if you will of word-to-token or vice versa (and I realize that may not be exactly the best analogy) was slightly offset.
Further, this sentence, "More technically, inference kernels produced incorrect results when used in certain GPU configurations." led me to understand how this could make it out into production - I'm sure ChatGPT has to be using tons of different GPUs in different configs to handle their volume, so it's understandable to me that their was a new matrix of config settings + code that made this bug show up.
I don't say any of the above to "excuse" OpenAI, but I also don't think they need any "excusing" to begin with. I don't think this was an unforgivable bug, and I appreciate them being so quick to explain what happened.
"More technically, inference kernels produced incorrect results when used in certain GPU configurations" has zero technical detail. The only information it is providing us is that the bug only showed up in some GPU configurations.
> Large language models (LLMs) like GPT operate using statistics to predict the next word in a sentence. Think of it like a highly sophisticated version of your phone's autocorrect that predicts what you're going to type next, but on a much larger scale.
> The system represents words as numbers, or "tokens," as part of its internal language. It uses these tokens to generate text by selecting them based on how likely they are to come after what's already been written.
> On February 20, 2024, an update meant to improve the system introduced a glitch in the part of the model that picks out these numbers. Imagine trying to grab a red marble from a bag without looking and accidentally grabbing a blue one instead—similarly, the model grabbed the wrong tokens.
> Because of this, the system started spewing out words that didn't make sense together. It's like if your autocorrect suddenly started inserting random, incorrect words into your sentences.
> The technical side of the bug involved something called "inference kernels," which are part of the system's operations. When used with certain types of graphics processing units (GPUs)—special hardware to process data—the kernels didn't work properly.
> Once the error was spotted, a correction was made to the system. After the fix, everything went back to normal, and the model resumed generating coherent responses.
It would be better if they elaborated on what "certain GPU configurations" meant because that's basically the central piece here.
It usually gets downvoted for banal aspects, but ones I've come to agree are important:
- oftentimes it's a lightly edited copy-and-paste of an attempt to summarize an article.
- even with edits, they're extremely long (this is edited, and its 250 words, about 1 page and 1/2 my browser viewport at 4K)
- usually off-topic because it's too broadly on-topic, i.e. its a summary of the article - ex. here, it isn't germane to the comment it's replying to other than 'if you want more info from them, ask what GPUs' -- it's unlikely the commenter needed the whole article ELI5'd to them in reply to their observation they'd like more info
Sort of "grey goo" for conversation, even with best intentions and editing applied.
> On February 20, 2024, an optimization to the user experience
At that point, about 10 words in, I already wanted to stop reading because it starts with the "we only wanted the best for our customers" bullshit newspeak. Anyone else going off on that stuff too? I'm pretty much already conditioned to expect whatever company is messaging me that way to take away some feature, increase pricing, or otherwise piss me off. In that case it was "not give any interesting detail at all".
Do you think OpenAI’s superalignment will ever allow you to make such a model?
Every time I ask a question, the answer could be 3x smaller. Even if it's right and not a hallucination.
Try:
> Act as a professional skeptic. Assess the passage for any deceptive, manipulative, or dishonest elements, including instances of doublespeak, inconsistency, fraud, disingenuity, deception, or sophistry:
It works on ChatGPT 3.5 as long as you're not trying to ask it questions about the Talmud. That will quickly devolve into handwavy "well, it's historically complex and difficult to understand" bullshit.
> ... > In conclusion, while the passage does provide a basic account of the incident, its lack of detail, reliance on jargon, and vague explanations leave room for skepticism regarding the severity of the issue, the transparency of the response, and the effectiveness of preventative measures. A more thorough and less technical explanation could have mitigated some of these concerns.
Cloudflare's postmortem: (https://blog.cloudflare.com/cloudflare-outage-on-june-21-202...)
> In summary, the passage does not exhibit overt deception, manipulation, or dishonesty. Cloudflare's statement seems to be a forthright attempt to explain what happened, accept responsibility, and outline steps to prevent future incidents. The inclusion of technical details and the acknowledgment of their error reflect a commitment to transparency, even if the promotional content at the end might seem slightly incongruent with the preceding apology and explanation.
Now the thing is to make a fast and reliable plugin that automatically puts the warning on the webpage and shame these corporates
In a medical setting this is equivalent to saying "the patient was coughing up blood, and now they're dead". It's relaying some potentially useful info, and I guess it's _technically_ a post-mortem analysis but it doesn't fit expectations for the phrase!
It looks like it was a mandelbug, which is hard to catch in a test environment.
~ "We tried a new optimization technique that modifies how next token candidates are chosen and there was a bug."
That would definitely produce the behavior people saw and makes sense.
What was the wrong assumption that the code made that caused this wrong behavior? Why was it not caught in the many layers of automated testing before it made its way to production? What process and procedural changes are being implemented to reduce the risk of this class of bug happening again?
Presumably all of that is playing out internally, but if the public postmortem is meant to instill confidence, you have to actually share some of the details, or else it becomes meaningless.