My take on an ai-agent powered research service. You configure the topics and the sources, then AI goes and does the research for you and distills the knowledge into bite size chunks.
Too cheap to be that specific I reckon. How many people will want to specifically research " tech & buy-side finance. " and when you get them you only get $120/year from them.
that's good feedback... i'm working on offering chat & full api access experiences too, but for a lot more $ :)
it works for any news sources, so in retrospect my hero isn't great. i'll update
Great, good update! Another question - are those 5 things on the right the only things it can research, or just some of the things?
I would hope something like this can do web requests and basic parsing of information on web pages, so I can send it off to research any topic and have a good chance. Is this the case? Or is it geared towards specific things?
those are just some examples of things i have used it for! when you login, you can fully configure the topics and the data sources that each agent has access to. so you're spot on that it does a pretty good job of navigating the web if you just give it google access.
the other sources in addition to google that you can choose to let it access are:
News API
arXiv
Hacker News
Financial Markets News
Worth making that clear. Even making the first or second agent say "Google" with a tagline "For any topic at all" would make it clear that this is an option.
I signed up and found that the agent is schedulable. That is worth pointing out in the feature list as I can see that being pretty cool for some people. For my case I am more of a one off research. So being able to turn that feature off might be good.
You may be in a space where each person like me has wildly different expectations though. Since "Research for you" is a very broad set of expectations. As someone who has played with things like autogpt, my expectations are low. No offence! It is just that getting AI to do this kind of stuff in a general case is pretty hard. At least this year lol! Next year it might be easy.
really appreciate the feedback. yes, the primary delivery right now is daily e-mail digest, so i'll highlight that better on the site, but sounds like your use case is more ad-hoc which would work nicely with chat.
definitely agree re: autogpt... getting it to actually do something was nearly impossible. i've found that by focusing on a more specific outcome, the agent approach works a lot better. with this it's been really cool to see how well it chases down internet rabbit holes to keep acquiring knowledge (well really just tokens) about a specific topic.
Yes for sure. Here’s an excerpt of mine from today. For this one the topic is GPT-4 and the sources are arXiv, Google, and HN
AI Generated News Summary for September 24, 2023
GPT-4
GPT-4, the fourth iteration of the large language model developed by OpenAI, has been the focus of recent research and discussions. One research paper titled 'LatEval: An Interactive LLMs Evaluation Benchmark with Incomplete Information from Lateral Thinking Puzzles' proposes a novel evaluation benchmark called LatEval to assess the lateral thinking abilities of language models. The study finds that even the advanced model GPT-4 struggles with lateral thinking and lags behind humans in this aspect. The benchmark aims to challenge language models with highly challenging tasks that are crucial for effective AI assistants. [Read more](http://arxiv.org/abs/2308.10855v2)
Another research paper titled 'Faithful Chain-of-Thought Reasoning' introduces the Faithful CoT framework for reasoning tasks. This framework combines a language model (LM) and a deterministic solver to ensure the faithful explanation of the reasoning process. The study shows that Faithful CoT outperforms standard Chain-of-Thought prompting on various benchmarks, including math word problems, planning, and question answering. The research also demonstrates the strong synergy between faithfulness and accuracy when using GPT-4 and Codex. [Read more](http://arxiv.org/abs/2301.13379v3)
SquadGPT is a chat interface to large language models (LLMs) developed by Sai_. It allows multiple people to participate in a chat with the LLM, providing a collaborative environment. The motivation behind SquadGPT was to address the challenge of limited access to commercial LLMs like ChatGPT Plus. By building a chat interface to the GPT-4 API, SquadGPT enables teams to leverage the LLM's capabilities without the need for individual accounts. The tool offers benefits such as enhanced visibility into AI costs, retained access to chat sessions, and better insights into users' guidance needs. [Read more](https://squadgpt.ai)
