Next decade is going to be fun.
Next decade is going to be fun.
Slack AI: Since I am a language learning model, my knowledge may not be current. Based on what I know from before my last learning cutoff date of March 23 2023, yes, Greg the senior developer is banging Becky from HR. They have been sending each other private messages every day for quite some time and are often offline at the same times throughout the day. According to their correspondence, you may be able to find them under the southwest stairwell or the janitor closet during the late afternoon. Remember, it would be ethical to approach them and talk to them earnestly about how you feel about their activities, and not to spread rumors. If you believe their knoodling is a violation of company policy, report them to the head of HR or your manager.
For the time being humans will probably still want to approve those tickets and PRs, but I think in the future in would be quite unreasonable for a human to attempt to critic an AI. It would be like someone as incapable as Magnus Carlson critiquing StockFish's Chess moves. Whenever an AI has reached human-level competency in the past they've always exceeded it just a few years after. And GPT-4 seems to be nearing human-level today.
Systems which can produce code unaided are obviously dangerous to humanity, but this isn't the concern of any single company. Businesses will do whatever makes sense for their business.
If you're talking about safety in the context of cyber security, then where the stakes are high (say in the development of banking software) you will probably see slower adoption of automation, but even there it will quickly become absurd to believe that humans are more capable at producing reliable and safe software.
No human can be have a real-time database of every known cyber security venerability in their head for example, but an LLM could.
I'd argue GPT-4 already has super human coding knowledge, it's really just on the intelligence side of things (reasoning, planning, logical consistency) that they're lacking a bit.
The next decade is going to be a security disaster, and nobody in the industry knows how to fix it, and the AI-agent proponents are basically sticking their heads in the sand over this because "there might be a fundamental security flaw in LLMs that makes this impossible to do" is just not a "fun" conversation to have.
Downstream of the parent comment we immediately have someone saying, "human oversight of this will of course be temporary until the AI gets smart enough." There's such a lack of perspective in the current computing industry of the risks involved here.
I'm starting to feel like the only way this narrative is going to change is for a company to actually get bitten by it, badly, and to have a bunch of internal secrets exfiltrated by a malicious email in a very public way. And even then, maybe the hype is too much, I don't know.
The same applies today, you just have to utilize any other internal attack vector, such as code injection in CI, on the deployment server, through a personal dependency etc.
None of the mitigation techniques for code injection work for LLMs.
Very literally, nobody in the industry has come up with a solution for malicious input into LLMs. It's an entirely different class of vulnerability. There isn't any dependency you can import or sanitization you can do or technique for getting rid of this.
In fact, I've seen research papers where security researchers are starting to hint that hardening prompts might actually be impossible for the current batch of LLMs.
I realise this is completely permited by Slack/other services and you agree to it as part of ToS, but people did with Facebook too. Doesn't lessen the rude awakening that awaits the majority of users.
Are you ready for EXPONENTIAL PRODUCTIVITY gains, optimise EVERY FACET of your life with generative AI that will commoditise your most personal and internal nature.
“Manna told employees what to do simply by talking to them. Employees each put on a headset when they punched in. Manna had a voice synthesizer, and with its synthesized voice Manna told everyone exactly what to do through their headsets.”
But the main thing I’d say here is that, at least in most companies, the things you write in emails or on the corporate wiki or whatever are not private to you – they are owned by the company. That doesn’t mean people don’t sometimes use corporate email for various private things, and it will certainly be used for sensitive things, eg it would be bad if something had read all the emails and could then answer questions like ‘is the firm thinking of firing X?’. But fundamentally, I think AI applied to corporate information retrieval is pretty different to whatever is meant by ‘next generation of Cambridge Analytica’.
There’s a technology in HBO’s Avenue 5 that I think is prescient: the rich owner of the space ship is pissed off that he can’t talk to people on Earth in real time due to the delay, since they’re pretty far from Earth. In order to appease him, they start using an AI that predicts what the other person would say and “simulates” the conversation, eliminating the delay. It’s a silly plot device used to simplify the script but I feel that’s where we’re headed all in the name of productivity.
Is your manager out of office? No problem! The AI will take over their slack and simulate them with Genuine People Personality (tm).
Pretty soon you’ll come back from vacation to find that your AI doppelgänger was prompt injected and you’re now responsible for the donuts every morning along with your paycut
This blows my mind a bit...I mean, what if we get stuck in endless loops of doing stuff because the AI suggested it? Imagine a "reply for me" mdoe in slack where the model does the things you listed...but what if the original message came from someone else's model? Then where did they get their instructions from? You can imagine some insane far future where product features are getting launched, blog posts written, services deployed and no-one is quite sure who is responsible for it all.
Likewise, iOS devices do analysis on all the photos on your device, matching up faces to names, classifying images, generating cute overview clips, etc.
It's not the AI you're thinking of, but it is analyzing your Stuff and applying intelligence to it. I don't see the GPTs of today as a revolution, but an evolution of the intent that has been there for two decades (and probably more).
It's not going to be very fun for people living under repressive regimes.
That same technology has the potential to make the haystack disappear, revealing just the needles.
Something like social credit becomes a lot scarier when the capability exists to feed an automated analysis of the sum total of a person's communications into it.
It's definitely going to be fun.
Why not just have a GPT model instead of the mailserver, just ask it about your new e-mails or ask it to tell something to your colleague when something happens. Have async meetings with the AI and just ask it for the meeting results based on how many participants took part in the meeting so far.
These ads are going to be so perfectly crafted the average human will have absolutely no resistance. It's going to be dark.
Not if we go back to smaller mail providers, which we really, really should.