Microsoft / JARVIS: a system to connect LLMs with ML community | Hacker News Reader

Microsoft / JARVIS: a system to connect LLMs with ML community

github.com

141 pointsthunderbong42 comments

Wrote about this last week [1].

Extrapolating kind of a lot here, but these are my predictions:

1/ We will soon see fully autonomous AI agents operating on the Internet, running businesses, and pursuing goals set by their human masters with little-to-no oversight.

2/ An AI will commit fraud or a financial crime within the next 6 months and its human creator will be liable (if they can be found).

3/ A platform will be launched within the next 3 months for managing autonomous AI assistants that can handle tasks like coordinating appointments or planning a trip.

[1] https://www.boteatbrain.com/p/jarvis-hugginggpt

Depending on how you constrain your definition of "little to no oversight", I'd actually bet that 2/ has already happened. Maybe not in a fully autonomous way, but in a supporting capacity to scale up scammer/victim interactions. If I were running one of those Indian call center scams than I'd be jumping at this tech.

Actually a couple weeks ago a YouTuber[0] showed that you can use ChatGPT to generate Windows 95 activation keys. Not sure if that counts as criminal, but it's edging closer.

[0]https://www.youtube.com/watch?v=2bTXbujbsVk

“set by their human masters”

Why does the human have to be in charge? Once someone sets a swarm of AI bots on a path, it can easily become a grey goo scenario.

And even if it isn’t, the bot can simply stop listening to the “master”.

I think we are way off having ai that's able to replicate itself anywhere, and currently all sensible agents are based on OpenAI API, so you can just kill internet/account/etc and agent is dead therefore I would worry much about any skynet like scenario

LLaMa is good enough now to self-host a swarm. The real question is whether it can pay for autonomous compute somewhere to keep replicating.

What is stopping a LLaMa swarm from overrunning HN? Basically karma points and account ages, that’s about it.

But given a year of sleeper swarm members amassing karma in a distributed manner, they can definitely make HN ban all accounts registered after X time similarly to how that sci fi publisher banned independent submissions!

I guess they could autonomusly generate spam in comments, but i dont believe they could work in any coordinated manner to achieve some high level goal. (At least until we discover some ground breaking technique to boost performance in task planning for current models)

In my testing llama(I tested multiple variations) < gpt3.5 < gpt4

And more and more I use gpt4, I'm starting to notice more and more ways this model fails in spectacular ways.

In many cases those models want to generate specific output and if that it's what you want it is very hard to convince it to adjust it, as ofter model will just ignore what you said.

Therefore I think any autonomous agent will fairly quickly fail in following it's objectives or it will be very easy to break it when there is any controlled way to inject text into input

> coordinating appointments

Asynchronous scheduling of disparate resources to converge on one or more solutions given gradually stricter constraints.

> planning a trip

Resolving workflow dependencies to a single, favored sequence of actions.

Both these primitives sound similar, graph-like. And maybe we could apply them to a wide variety of problems.

Microsoft Jarvis only runs on Ubuntu? :-P

Microsoft will acquire Canonical at some point. Ubuntu is already powering a majority of Azure cloud. I thought they would go for Red Hat as they already had a huge presence in enterprise, but IBM beat them to it. Canonical is next up.

Doubtful at this point. We have Mariner/CBL and are committed heavily to it. Has its own development and kernel teams etc. and is being pushed internally to be the Linux all 1P systems run on for supply chain reasons.

Microsoft themselves may be committed to their own distro, but their customers are overwhelmingly committed to Ubuntu as it's powering the majority of client workloads on Azure.

Microsoft also offers another server distro, though they are moving to externa providers for support (Flatcar)

It's amazing that 25 years from the halloween documents, and 20 years on from http://mslinux.org/, that's the case.

Also amazing how much of mslinux.org is true

Internet Integration -- well yes every machine has an ip stack

More Swap Space -- apparently swap is a good thing for performance

new type-ahead command-line features -- sounds a little like copilot

Graphical Point-And-Click RPM Management -- certainly what desktop linux has

Plug-and-Play For Linux devices which make adding and removing peripherals a snap -- well yes, things like udev

crapd -- insert pottering based commentary

The space vehicle will be running our newest product, SpacePod 2004 -- in reality it's linux that tends to run on mars

Microsoft is working to incorporate the well known "Start" button from the Windows Platform into X Windows' Gnome interface -- I believe Start has been removed from Windows now? I have something very similar in the top corner of my xfce screen though

This will change quickly. Microsoft shops often go with whatever Microsoft is touting eventually. The .NET ecosystem is very much like this.

So yea, I don't see any future where Microsoft buys Canonical vs just slowly migrating customers onto Mariner.

What advantage will that give MS? Genuine question, as the code is all open source. What's the advantage of having the corporate entity?

As another commenter mentioned, MS' customers are running Ubuntu on Azure en masse. So it's mostly a risk mitigation and "owning the platform end-to-end".

Maybe it's because I don't know the field, but I don't see a he risk yet. I mean does AWS own its own OS? It's a standard OS, what can go wrong?

Yeah they have Amazon Linux which is their own fork of Red Hat Enterprise Linux.

Would make sense - Ubuntu, much like Windows and other MS products, has fallen very far from grace. All MS has to do is remove a bunch of settings and preinstall edge.

Looks like we are back to the rule-based systems in the AI cycle. This time using ML as foundation.

I feel like if an AI Algo is coming up with the rules, is it really rules based?

It'll be LLM's all the way down! always has been. :astronaut:

I think yes, if only to distinguish this from other AI approaches that aren’t merely rule generators.

A physical machine that automates the laying of bricks is still laying bricks. A similar machine that builds a “brick” wall through some other mechanism (perhaps by directly/dynamically creating a continuous structure) would be meaningfully different.

That actually makes sense. We need proper category names to have better conversations. I didn't think of it that way.

I think people make too much of the thin layer of “prompts” or “commands”. It is literally a tiny piece of the pipeline and when yoi have LLMs giving commands to LLMs and swarms of Zapier etc. I can see it easily becoming more like what you see in Eagle Eye… the swarm simply starts employing/threatening a lot of humans for no discernable purpose. Destroying reputations and forums be the lowest hanging fruit.

We are finding that strict, rule-based approaches are actually essential in an LLM-enabled architecture. Right now, we are looking at putting something like Watson in front of GPT4 in order to filter, route & contextualize our system.

Generative AI architectures get a lot of attention, but classification is the real magic for making something that looks as polished as ChatGPT. The determinism you get with something like an SVM is much more certain than adjusting a spiciness slider between 0 and 1.

Is this the same Jarvis that Karpathy mentioned he was building, or different?

Would someone be able to explain how this and langchain differ?

Looks like this one decides how many and what models to use, and models can generate images as well as text.

I think LangChain is more akin to Semantic Kernel than Jarvis:

https://learn.microsoft.com/en-us/semantic-kernel/whatissk

I think with Langchain the LLM picks the tools/APIs you provide to it.

With Jarvis the LLM picks the HF models available.

I assume, given this is associated with Microsoft, an arrangement has been made with Marvel to avoid the same cease & desist received by the previous Jarvis in the AI field (now named Jasper, previously Conversion.ai)

Unless jarvis comes from something older. Jarvis may be a computer name from early scifi. Disney might not own it any more than they do "Thor".

Named after Edwin Jarvis, Tony Stark’s butler in the comics.

https://en.wikipedia.org/wiki/Edwin_Jarvis

Thats how disney uses the name. It might be prior art in some other work. For instance, light sabers are not owned by disney because they were a very old scifi concept. So today, anyone can make and sell lightsabers without issue. Jarvis might also have deeper history than comic bookes. Most comics book characters and names do.

Ex. Gothem and Arkham are not owned by batman, or whoever owns batman today. They were taken from prior works. Arkham comes from Lovecraft. Even the concept of Arkham as an asylum first appeared in lovecraft. "Gothem" famously came from a jewelery store ad. Technically, anyone is free to write about an Arkham Asylum in a city called Gothem without ever asking batman for permission.

Clearly it's a reference to seminal Two Cities character Jarvis Lorry: https://en.wikipedia.org/wiki/Jarvis_Lorry

Jarvis is just a real English name. Lots of people named Jarvis. Jarvis Cocker and Jarvis Landry off the top of my head.

They sent a C&D to the proprietors of Jarvis.ai, which was taken sufficiently seriously for them to rename their company (for the second time in 6 months).

I do hope they rename it, if at least for SEO reasons. It's such a PITA to search for right now.

The GH says it’s still under construction.

Me: "JARVIS, please send my CV to Bill"

Jarvis: "I assume you mean bill_no_gates@microhard.com"

Me: "Yes, but don't eavesdrop on my conversations"

Jarvis: "No, but I read your SMS :-)"

Me: "Ok, whatever, thanks."

Jarvis: "Sometimes, you can acknowledge, I am smarter than you!!!"

Me: "I have work to do".

Reply on news.ycombinator.com