willahmad

This can happen because of training data. Imagine you have thousands of legal documents rejecting things to Iran.

eventually, model generalizes it and rejects other topics

It can happen because training data contains lots of rejections to groups (Iran sanctioned, don't do business with Iran and so on). Then model might be generalizing 'rejection' to other types of responses

"I will read it after this 15 min short video"

you are right, it probably won't fix the traffic, but it doesn't matter because if you are not steering the wheel then you can work or study while in the traffic.

there is a huge economic impact

it's not only about number of users. You also have regulations in different cities and states, existing models trained based on past supply demand behavior, somewhat optimized workflows for the ride hailing industry, payments, risks.

They both can build themselves, but if you provide solely self-driving ride hailing, a lot of times customers might not be able to find cars in upcoming 6-7 years until they ramp up full production to meet demand.

moat is in institutional knowledge of operations in ride hailing industry.

Also, pleasing the customer, imagine opening Waymo app and not being able to order a taxi 40% of the time. With Lyft/Uber you can easily switch the ride mode and get a car with driver if all self driving cars are busy

From market standpoint I am curious if Google or Amazon is willing to acquire Lyft for the tech and customer base.

Lyft is a good distribution channel for their self driving car initiatives with good coverage, Uber is too expensive at this moment.

Here's my experience with these tools:

Good: I can prototype things very quickly thanks to these tools

Bad: After couple of vibe coding iterations, I don't have a mental model of the project.

Good: When I open my past projects where I have very good mental models, I can come up with a nice prompt and build anything quickly again.

Bad: After couple of iterations I become lazy, and eventually my mental models break.

There's definitely a use for these tools. But be careful, job of engineers are not only coding but also training their memory to build solutions and bridge real world problem with software solution. If you lose this skill of thinking, you will be obsolete quickly

I think what they are trying to do is a type of forward deployed engineer, but with no employment.

More they train such engineers more profitable for them to spread the word.

In 1st cohort, they're probably going to accept extrovert people with active social presence.

It could have been, but you still need to compress and transform OpenAPI, here are some examples:

* Stripe OpenAPI has ~300 endpoints, imagine adding it to 128k context window, you will run out of context

* Most APIs return unix timestamp, which LLMs don't understand and can't convert to human readable format without additional tools

* With OpenAPI you are trying to be extensive (see Stripe API, contains everything), with LLMs you want to provide basic functionality (only transaction related functionality for humans, even that is limited to last month or week)

With MCP, I think authors envisioned long running processes with duplex communication (e.g. notification about progress).

You can obviously achieve all these with OpenAPI, but it requires some engineering effort to adjust for LLMs. And when people feed OpenAPI as it is, they would start complaining about sub-par quality, because it didn't understand the timestamp

I wish we get better alternatives to Anthropic sooner. Fortunately, OSS models like GLM, Qwen are catching up

Obviously, it's good for them if things are regulated, but bad for all of us.

First of all, congratulations

Honest question expecting your honest response. How much of it's written by AI? Recently, I noticed I am liking thoughts of real people over the super structured, emdash-y, zero grammar mistake AI texts.

I want to support authors who share their experiences, rather than guiding AI to write thoughts

But they provide sources/links, which I click occasionally and they have utm source with chatgpt

Depends on the questions.

If a technical problem, before my questions were leading either to official doc or stackoverflow, now I get possible solutions to try

Other types of questions like reviews, product comparisons GPT shows couple of relevant links to reddit and I go check them to see if summary was right and then surf relevant subreddit.

If I assign a number, I would say 80% of my requests lead to answers inside the GPT, other 20% lead to links

From a personal experience, at least 50% of my Google searches moved to ChatGPT and Claude. Since, I saw many similar transitions with my friends and coworkers, I was expecting higher numbers for OpenAI.

Do you see a similar transition in your network?

I have built similar gateway but as a desktop application which runs locally and gives you a lot of control over enabled MCP tools and allows you to have a separate profiles for different combinations of MCP servers and credentials (e.g. profile with production API keys, sandbox keys and etc,.).

give it a try here: https://getaiko.app/downloads

Trust takes time to build and only a moment to lose. That’s why we're trying to build the trust by committing to transparency and security.

Right now, we’re undergoing the CASA Tier 2 security assessment for our desktop application. This process, required by Google for publishing our Gmail and Google Calendar integrations, which reassures that we are not transferring data to our servers and storing it securely on your local machine. We also planned to obtain SOC 2 after we validate the demand for the tool.

We also encourage our users to inspect the network traffic from our application to verify this for themselves.

I have built bunch of MCP servers so far, and here is my take (and why I love it)

There are 2 types of actions, (1) read-only and (2) MCP servers/tools with side effects

1. Read-only types. As you know, LLMs are static models, they don't learn things between your sessions, hence you need RAG or better prompting to enrich the context to get the best results for your use case. MCP helps you with enriching that context. For example, you want to categorize your recent emails, you can obviously copy paste all of them as a context, then ask LLM to categorize. Or you use MCP server to pull emails and add it to the context. This is very similar to RAG, but heavily personalized for your use case (e.g. by pulling data from Jira, Github only when you ask it)

2. MCPs with side effects (write/delete) - Here you can leverage NLP capabilities of MCP to take actions. For example, send email or create an event in your calendar, or enrich the contents of Jira ticket. Same as (1), you can ask LLM to come up with contents and then manually copy/paste it to your calendar to create an event or leverage the convenience of MCP tools to do it automatically.

Here are some demos from my MCP servers:

* Integration with Google Calendar to create events based on information of another MCP server (memory) - https://www.youtube.com/watch?v=ZgEy6Y1kfn4

* Here you can see how easily you can integrate your OpenAPI based spec and use human language to query it - https://x.com/getaikoapp/status/1945278307496235482

I am working on easiest way to manage all your MCP configurations and credentials. Additionally, I have added couple of interesting MCP servers, like converting any OpenAPI v3 spec to MCP server (including support of authentication): https://x.com/getaikoapp/status/1945278307496235482

After working on and using many MCP servers, I hit couple of issues multiple times:

* Do I configure 2 MCP servers of same type for 2 different API Keys or do I manually update configurations all the time? (e.g. production and development environments)

* when I have too many tools enabled, I noticed that either I am hitting context limit too quickly or LLM is hallucinating when choosing a right tool

* Some MCP servers expose a lot of tools, I want to disable some of them forever, instead of doing configuration per AI assistant (first for Claude, then Cursor and so on)

* Most MCP servers are hosted by third parties, as a privacy conscious person, I do not want to share my credentials with third parties.

And I am building Aiko - AI tools marketplace: https://getaiko.app

NOTE: Gmail and Calendar apps are currently under CASA Tier 2 security assessment, hence not published to production. But you can see demo usage here: https://www.youtube.com/watch?v=ZgEy6Y1kfn4

Would you prefer a local only tool?

As a privacy conscious person, I am developing a desktop/local only app and collecting feedback from beta testers. I would appreciate if you can participate as a beta tester and share your feedback.

At the moment, app is only available for MacOS, and in the process of security audit (not SOC-2 yet, but we will get there). I am planning to sell a license.

Please let me know how to contact you, if you prefer to send email here is my disposable email: kzphefll8@mozmail.com (to prevent spam emails, not sharing my original email)

Thanks for taking time to explain this.

> But, I suspect, if the model is able to handle language at all, you'll always be able to get a representation of the prompt out in a text form

if I understand correctly, system prompt "tries" to give a higher weight to some tensors/layers using a text representation. (using word "tries", because not always model adheres to it strictly)

would it be possible to do same, but with some kind of "formulas", which increases the formula/prompt adherence? (if you can share keywords for me to search and read relevant papers, that would be also a great help)

I'm not a ML engineer and only have surface level knowledge of models, but I’ve been wondering, would it be possible to train models in a way to be able to embed a system prompt in a non-textual format?

Ideally, something that’s lightweight (like cheaper than fine-tuning) and also harder to manipulate using regular text prompts?

> This is why we assign an engineer to review every PR of a vibe coded project and they don’t really need all of the context to detect bs approaches that will surely fail.

I see this trend in many companies as well, just curious, how do you make sure engineering time is not wasted reviewing so many PRs? Because, some of them will be good, you only need couple of your bets to take off, some definitely bad

B2C might not be sticky, but enterprises can't easily migrate and it takes a quite long cycle of decision making to replace one product with another, in this case LLM.

I think everyone is betting on making a good product and convince enterprises to buy from them, this is where you can get sticky customers

UI looks really neat and pleasant to use. Does it create a todo list per prompt similar to Claude Code?