The GPT-V accepts screenshots of your desktop and application GUI as input. Please ensure that no sensitive or confidential information is visible or captured during the execution process.
This project is the future that we were promised but it is under threat by supposed legal challenges.
What if the demo asked the email to be sent as a message via whatsapp desktop instead? That would (according to their anti-freedom lawyers) constitute an offense worthy of legal threats.
The tech industry needs to reckon with ToS trolls before it's too late.
Technically speaking though, user facing websites are built with quite weak UI accessibility in mind. Partly to prevent others bots from using it. I worry less about the LLM/AI particularities, and more about adding yet another comprehensive layer to the stack. It’d not exactly be standing on solid foundation.
During the 2010s Web 2.0 there was a brief moment where open APIs and such were trending, which was a bit of rejuvenation of interoperability across companies, domains, applications etc. To simplify we can call it cross-app interactions, like auto hotkey, Automator, etc. An LLM controlled broker falls in that domain as well. But now that “closed binary fuck you” model came back industry wide, it’s almost impossible to build such things. (It’s hard enough to build integrations when people are cooperating. When they’re actively adversarial, things generally break quickly, if they work at all)
Adversarial interop should be a digital human right.
Put an API layer in front of UFO and all of a sudden we're one step closer to the unshittification of our digital lives.
Never once was the conscious decision made to not be accessible to prevent bots from scraping.
This is not hyperbole, that's what they will do based on precedent. And it doesn't matter about the validity of their claims because the calculation as a victim to these legal threats is that this $800bn megacorp is able to ruin your life for what amounts to less than pocket change and these big law firms are incentivized to come after you.
Adversarial interop should be an inalienable digital human right. This way companies will be forced to give API access or risk interoperability being legally scraped against their will.
https://docs.google.com/document/d/14p_iPhIKjDoTGa2Zr_5gPV9_...
Some specific errors I did notice:
- In the section with a processor unit subfactory, "The specific items being manufactured are not directly visible because the UI for the assembling machines is not expanded to show their recipes" is false, the player has pressed 'Alt' and the items being manufactured are shown. So this part of the response is plain wrong.
- " There are four distinct colors of science packs visible on the conveyor belts: red, green, blue, and purple, corresponding to the various levels of research complexity in the game." Only red/green/blue are shown. ChatGPT didn't make additional references to purple science but it was odd that it mentioned them at all.
- "In this Factorio image, we see a railway intersection that includes train signaling and a train crash" this is a deadlock, nothing has actually crashed. This is a minor nitpick but ChatGPT repeats the error throughout the analysis. And it also suggests that ChatGPT might not fully grok the "race condition" side of train scheduling, since deadlocks occur specifically because you're trying to avoid collisions.
I didn't want to spend too much brain calories reading all 20 pages in depth :) My general conclusion is that it's not useful enough for GPT-assisted Factorio play, and too flaky for any sort of automation of trivial Factorio tasks. I think it's plausible to make a FactorioGPT, but I doubt OpenAI's pretraining and RLHF resources covered this specific niche.
On the other hand, I shudder to think of the millions of man hours required to arrive at this solution, when simple UI guidelines, or better yet, an API, would have solved my problem far more simply and efficiently.
They are on a roll lately, and seem to have beaten OpenAI to GPT-Agents with this release.
OpenAI gave Microsoft the model weights, and Microsoft hosts it on Azure for MS Research. None of the Azure usage analytics goes back to OpenAI except via bug reports and publications.
It said it could not do say add an image to a ppt/create one. When it clearly could. Chat was was overly simplified and fixed the wrong problem. Randomly changed say Excel sheet with no clear undo after multiple steps Plus how it works is hidden, so not sure if they are using 3.5, 4 or something else. So no idea if that is causing the problem.
MS is putting out a lot of things quickly. But the quality is just not there in my experience. They are doing way too many things too fast to make any one thing good.
Sometimes these are automated with ”robotic procesd automation” tools. Something like UFO could streamline the process.
On a side note, it looks like this thing can be a terrific cheating tool in Strategy video games..
CogVLM: Visual Expert for Pretrained Language Models
CogAgent: A Visual Language Model for GUI Agents
https://arxiv.org/abs/2312.08914
https://github.com/THUDM/CogVLM
https://arxiv.org/pdf/2312.08914.pdf
CogAgent: A Visual Language Model for GUI Agents
Abstract
People are spending an enormous amount of time on digital devices through graphical user interfaces (GUIs), e.g., computer or smartphone screens. Large language models (LLMs) such as ChatGPT can assist people in tasks like writing emails, but struggle to understand and interact with GUIs, thus limiting their potential to increase automation levels. In this paper, we introduce CogAgent, an 18-billion-parameter visual language model (VLM) specializing in GUI understanding and navigation. By utilizing both low-resolution and high-resolution image encoders, CogAgent supports input at a resolution of 1120×1120, enabling it to recognize tiny page elements and text. As a generalist visual language model, CogAgent achieves the state of the art on five text-rich and four general VQA benchmarks, including VQAv2, OK-VQA, Text-VQA, ST-VQA, ChartQA, infoVQA, DocVQA, MM-Vet, and POPE. CogAgent, using only screenshots as input, outperforms LLM-based methods that consume extracted HTML text on both PC and Android GUI navigation tasks—Mind2Web and AITW, advancing the state of the art. The model and codes are available at https://github.com/THUDM/CogVLM .
1. Introduction
Autonomous agents in the digital world are ideal assistants that many modern people dream of. Picture this scenario: You type in a task description, then relax and enjoy a cup of coffee while watching tasks like booking tickets online, conducting web searches, managing files, and creating PowerPoint presentations get completed automatically.
Recently, the emergence of agents based on large language models (LLMs) is bringing us closer to this dream. For example, AutoGPT [33], a 150,000-star open-source project, leverages ChatGPT [29] to integrate language understanding with pre-defined actions like Google searches and local file operations. Researchers are also starting to develop agent-oriented LLMs [7, 42]. However, the potential of purely language-based agents is quite limited in realworld scenarios, as most applications interact with humans through Graphical User Interfaces (GUIs), which are characterized by the following perspectives:
• Standard APIs for interaction are often lacking.
• Important information including icons, images, diagrams, and spatial relations are difficult to directly convey in words.
• Even in text-rendered GUIs like web pages, elements like canvas and iframe cannot be parsed to grasp their functionality via HTML.
Agents based on visual language models (VLMs) have the potential to overcome these limitations. Instead of relying exclusively on textual inputs such as HTML [28] or OCR results [31], VLM-based agents directly perceive visual GUI signals. Since GUIs are designed for human users, VLM-based agents can perform as effectively as humans, as long as the VLMs match human-level vision understanding. In addition, VLMs are also capable of skills such as extremely fast reading and programming that are usually beyond the reach of most human users, extending the potential of VLM-based agents. A few prior studies utilized visual features merely as auxiliaries in specific scenarios. e.g. WebShop [39] which employs visual features primarily for object recognition purposes. With the rapid development of VLM, can we naturally achieve universality on GUIs by relying solely on visual inputs?
In this work, we present CogAgent, a visual language foundation model specializing in GUI understanding and planning while maintaining a strong ability for general cross-modality tasks. By building upon CogVLM [38]—a recent open-source VLM, CogAgent tackles the following challenges for building GUI agents: [...]
https://github.com/microsoft/UFO/blob/main/ufo/llm/llm_call....
that app that like the entire united states uses for pc work every day?
i still cant copy paste a code block, or copy paste literally anything. i think microsoft should use AI to learn how to code code blocks in chat or they should ask chatgpt how to use the clipboard of their own OS
Teams is actively developed along the lines that deliver greatest value to Microsoft at the expense of their customers (a business model increasingly popular these days):
1) implementing corporate-nerfed versions of vanity features introduced by competitors in group chat space (Slack, Discord);
2) broader integration with everything else in Microsoft's corporate ecosystem.
You can be excused for thinking Teams is just a crappy chat-based interface to SharePoint, because this is what it effectively is (Don't have SharePoint? Sucks to be you.).
Copy-paste? What are you? A corporate smartass? There's no budget left for smart-ass features - it's all in lock-in features, where the RoI is much greater.
When a game studio releases new game, looking at previous games they released is a solid indicator of what quality we can expect. Same goes for software companies making new piece of software. Combined with how LLMs seem super useful until I want them to do something original, I would be very sceptical even in case a corp like Apple came up with this announcement. Seeing it comes from Microsoft, not trusting it isn't result of "hating them is popular", its result of experience.