Theorem equiv_pq_qp : forall (p q : Prop), (p -> q) <-> (q -> p). Proof. intros p q. split. - intros p_imp_q q_imp_p. apply q_imp_p. apply p_imp_q. assumption. - intros q_imp_p p_imp_q. apply p_imp_q. apply q_imp_p. assumption. Qed.

... together with a lengthy and convincing explanation in natural language.

Sophists would be delighted by these mechanized post-truth AI systems.

I've been actively using Kagi[0] for the past few months and what really works brilliantly there is the notion of "lenses", AKA "per-site" search filters. So you can say - "boost reddit.com, suppress bullshitspamsite.com" and set this up for specific contexts - PHP dev, recipes, etc.

I see that Phind has some notion of this, so will have a play - but it's just this relatively simple "limit to X sites" option that I think makes a big difference in the usefulness of search for me.

[0] https://kagi.com/

"Best California style burrito in Austin".

Nearly every engine shows me burrito shops in California, some give shove reddit links to the top. Google was the only decent response. Phind response is what I would expect from an assistant who researched this for 5-10 minutes of searching the web. Great work!

(now add maps to those results!)

Absolutely fantastic stuff, I’m excited to add this to my tool-belt. There’s a specific feeling of knowing that an answer to your question is very simple and exists somewhere on SO, but the mental effort of sifting pages of answers seems unappealing. It seems like Phind is well suited to do this job for you!

That simple act reduces your cognitive load, as a programmer you mind is two steps ahead assuming this trivial solution you know conceptually but do not want to spend the mental cycles and looking for the quick code so your mind can jump to the next step in the flow.

https://phind.com/search?q=How+should+I+filter+a+dictionary+...

https://phind.com/search?q=How+would+I+use+php+to+encode+wit...

https://phind.com/search?q=How+do+I+get+my+data+from+a+datab...

For more complicated prompts it misses the mark a bit but this may be going outside of the intended use case.

https://phind.com/search?q=Create+a+python+class+that+can+pa...

It would be nice if it annotated the projects it suggested in the response. The above query talks about a project called "mov" in the "provided code" but I do not see any code provided.

Running it again, I got a code example:

> Create a python class that can parse MP4 headers

> To create a Python class that can parse MP4 headers, one can use the pymp4 library available on GitHub.

The Box class from this library can be used to build and parse MP4 headers.

The following code shows an example of using the Box class to build an MP4 header and then parse it:

from pymp4.parser import Box from io import BytesIO

header = Box.build(dict( type=b"ftyp", major_brand="iso5", minor_version=1, compatible_brands=["iso5", "avc1"])) print(header)

parsed_header = Box.parse(header) print(parsed_header)

That question wasn't my first question.

For clarity I asked different questions in the top search box assuming that would kick off a new session.

The performance could be improved. I've having to wait several seconds before the summary is created.

https://phind.com/search?q=why+is+funcref+not+working+in+God...

Code snippets are fine, but I'd actually rather just use ChatGPT to generate code snippets and ideas, and have it explain to me details and other options in a conversational way.

Thus in general this approach, like most other AI search options, doesn't excite me.

At a high level, I think AI search is terribly hard to solve. You're fighting against existing user patterns heavily tuned to the nuanced behaviours of existing search. So you have to find a specific use case that Google sucks at (and while StackOverflow's fall from grace somewhat lends an opportunity, I think it's not sufficient for most developers to start looking for a different solution) and that you're 10x better at.

But 10x better means your speed has to be just as fast, your answers have to be significantly better, and crucially you have to be nearly 100% consistent, which is brutally hard for most AI approaches.

At that point I wouldn't even market yourself as search. In this context maybe something like "AI-powered software development assistant".

For example, "how to debug pthread race conditions in c++" is asking for a guide where no single comprehensive source exists. Phind combines information from Stack Overflow and other websites to generate a more comprehensive guide: https://phind.com/search?q=how+to+debug+pthread+race+conditi....

Generating guides that aren't simply addressed by docs or individual Stack Overflow answers are where Phind really shines.

This is made even worse by the verbosity of Phind's answer; it could've simply given me a sentence about tsan, mutexes and reverse debugging, but it decided on a highschool essay which took as much to read as it took to skim five different methods in some stackoverflow answers.

I really don't think it's there.

One of the things that comes to mind is searching for exact error messages or codes, but for that I want something more like grep rather than AI.

That's one of my pet peeves about DDG -- they're indescribably bad about that. I actually haven't tried the same query on Bing to see if it's a Bing index problem, or if DDG just isn't good about it

Cool. Conclusion is that I need to get better at writing queries:

I wanted an overview of Python APIs to choose from to play sound files, and importantly, I wanted to know how / if each API offered a non-blocking parameter. In my use-case I am playing sounds, after detecting some elements during video processing, and I find it blocks my realtime video stream UI.

My experience refining my query - I started with one query, and settled on an increased redundancy of my query:

[NOTE: I typo'd asynchronously as "asynchonously", unintentionally]

FIRST = how do i asynchonously play sounds in python (https://phind.com/search?q=how+do+i+asynchonously+play+sound...)

Result from PHIND for FIRST was promising enough for me to refine my query. I felt like the beginning portion of the answer it provided was useful. And I felt like the subsequent portion was not. It listed various Python API for playing sounds. And only for the first did it supply concurrency details.

SECOND TRY = how do i asynchonously play sounds in python, and what are the parameters for synchronous or blocking playback (https://phind.com/search?q=how+do+i+asynchonously+play+sound...)

Result for the SECOND was more loquacious, however it supplied more detail, and satisfyingly more concurrency configuration instructions for e.g. the "playsound" Python API.

So overall, kudos! And, I will use this product in the future.

> In TypeScript, how can I have a component restrict the types of its children to only certain types of components? I would like a static type error.

The correct answer is "this is impossible". Every AI I've tried this on hallucinates some nonsense code that doesn't actually work. Sadly Phind is the same. It says "here's an example of how to do this" with a code sample, and then links to a StackOverflow post saying it's not possible :)

I know quite a lot of people, if not allowed a google search and only get vscode to try something, that would take a stab at your typescript question and not be able to tell you that it is impossible.

Asking gpt to be a search engine is the same as asking a human currently really: I vaguely remember dates from WOII history, so when you ask me in a pub quiz, I will confidently tell you something and then where to find or look up the actual answer which might be slightly or completely different, depending on if I actually matched the right event with the right date and if I remember the day and month or not.

What are you referring to here? The title of this submission says "Phind.com - Generative AI search engine".

Once again proves that it is more untrustworthy than a normal search engine like Google. This AI hype of LLMs is truly going to subside very quickly given that the promises made by what ChatGPT, Bing AI, Bard and now Phind are collapsing right in front of us when tested.

The truth is, almost no-one here would trust their output, and now needs a extra review by a human for each output. Henceforth, there is no point in using it as a search engine or to generate anything factual given it is going to hallucinate nonsense like that and can be easily tricked to output incorrect answers, making it very unreliable.

Q: Is this general usability issue being discussed more, due to the growth of new product teams working with generative AIs?

Some discussion: I am drawn to wonder if there is an innate value to include permalinks (or semi-permanent) ... to save each query's results. Or perhaps do something better which I've not thought of yet.

The issue in more detail, (I'm naive on the subject so forgive the redundancy:) Results aren't highly deterministic - randomness and date-sensitive models seem to generate, probabilistically, different results for the same query. Especially for Generative AI products that exist as webapps, where hitting 'reload' could provide inferior results and simultaneously eliminates superior prior results.

Non-determinism is a 'surprise' quality which can enhance and degrade the value of a generative AI tool, in my experience when using this category of tools. It's fun (endorphin rush) to spin the Roulette wheel and query again. And it's painful when the prior results ranked More Useful in one's mind.

Ideally for usability's sake, these products would include some affordance to the user, to manage that endorphin / pain cycle.

[Note, A complication, here with PHIND being the context. PHIND's HTML renderer itself is being iterated on .. this is a product-launch after all. So, results also look different not for data/algorithm reasons.]

This could be realized as a web-browser feature - an enhanced History. The current "History" model of browsers is a trivial response to the REQUEST/RESPONSE HTTP mode. The HTTP spec (last time i read it was in 2005) leaned towards everything-is-deterministic except for time.

Example: I input this query https://phind.com/search?q=linq+query+to+filter+every+other+... and it listed one code snippet. I re-queried and received a different snippet. Honestly, I like both results. As of this moment I'm writing, I can only generate one of the two results.

Building on this, I see now permalinks could be challenging to integrate "perfectly" to PHIND's "Ask a followup question" value-add feature, because that feature depends upon the original query's model to provide "perfectly" (my word) useful "followup" results. And a permalink's model should be expected to quickly be stale, and quickly discarded, at this time.

I can conclude then, there is less value to permalinks at an app level, since the value-add features of the e.g. PHIND webapp may be compromised if they're intended to also work with the permalink data. And perhaps more value would comes from a trivial global, static history at the browser level instead of app.

> Based on the given context, there is no clear answer to the question "what's the best build system for Visual Studio Code on Windows". However, there are some relevant information that can be helpful in understanding the build system options available for Windows and Visual Studio Code.

> [... basic rundown on build systems and a list of the common suspects ...]

I'm impressed.

I prefer search engines to do their job and provide quality links, not lengthy passages about the subject I'm searching for. For simple answers it's fine, but search engines pretending to know all about a subject, is cringey. Scraping the web and mashing together what it finds.

Each website has its own way to present things, and in what order. When hoovered up, taken elsewhere and squeezed out through AI's front end, it comes out a tangled mess.

Just now I needed a refresher on the HTML table scope attribute, so I used Phind to look up html table scope. It returned a wall of text about tables. Not inaccurate, but a wall of text. It was easier to click on the w3schools links where it's nicely laid out with clear examples of scope, with clear information about related attributes and use.

Naturally, the first thing I searched for was something I’ve put on the internet that is not well indexed by Google.

When I search “how to find a lobby with rivet.gg” (that’s our most basic functionality), it doesn’t come up with an answer. However, it does spit back a descent summary of relevant features we provide.

What’s odd – it always continues on to talk about an unrelated company. The first time, it started talking about top.gg as a way to add Fortnite-style bots to your game. (They’re actually a Discord bot directory.) The second time, it mentioned a completely random Reddit called Guilty Gear Strive.

Can you elaborate on how it finds these related topics?

In the case of “how to find a lobby with rivet.gg”, it finds your docs and looks at https://docs.rivet.gg/docs/concepts/matchmaker/ but fails to extract how to put a player in a lobby. We'll take a look at this case in more detail. Our goal is to be great at parsing most/all docs.

I am curious about whether it hallucinates api methods that don't exist, like ChatGPT does. I haven't seen that yet but the underlying concept of an LLM is the same.

There is a small accessibility problem with the way it interacts with screen readers. it speaks duplicates of the text for some reason. I am using NVDA.

(a.) Generate a short summary to start with & expand using something like `show more` (b.) Include reference images if you can.

A very generic query I agree.

Response was

Reverse ETL is a process that involves copying data from a data warehouse to operational tools such as SaaS tools used for growth, marketing, sales, and support. Traditional ETL involves replicating data from a data source to a data warehouse. Reverse ETL is a relatively new category in the data space that enables real-time decision-making, increased collaboration between departments, and streamlined workflows. Reverse ETL allows different departments like product, marketing, sales, engineering, and customer support to be more self-sufficient and effective by giving them direct access to the data they need within the tools they use. Here are some key differences between ETL and Reverse ETL:

- ETL is a one-way door, whereas Reverse ETL is a two-way door. - ETL pipelines are used to read and write data from operational systems to a data warehouse, whereas Reverse ETL is all about syncing the data in your warehouse to your downstream business tools. - ETL is typically used for batch-oriented data integration, while Reverse ETL is usually used for real-time data activation. - ETL focuses on integrating data from multiple sources, and Reverse ETL is focused on making data available for consumption.

Reverse ETL requires writing reverse SQL, which can be challenging. Reverse ETL tools provide a visual interface to choose which query output columns are used to populate standard and custom fields in your operational system. Reverse ETL tools are generally more powerful and cost-effective than iPaaS and open-source solutions. Reverse ETL enables any team, from sales to product, to access the data they need within the systems they use. Reverse ETL can help business users get the data they need without relying on engineering resources, making it easier for analysts to focus on high-level tasks. Reverse ETL can also improve the quality of your data by catching errors and inconsistencies in near-real-time and fixing them before they cause problems. Reverse ETL is a compelling process with many benefits, but it does have limitations. It does not provide an interface to activate your data into personalized, cross-channel customer experiences. This gap is where a customer data platform (CDP) comes into play. A CDP unifies all your customer data and consolidates it into a single profile view for each customer. Reverse ETL is a relatively new technology that sets businesses up to do more complex operational analytics than they could with BI tools alone. It has been around since 2018, but it didn’t really start trending until 2021. Reverse ETL is a technology for taking cleaned and processed data from the data warehouse and ingesting it back into business applications such as Salesforce, where it can be used for business operations and forecasting.

Coz, I am sure the response generated would definitely not be the same if I use same question twice. So, that would help.

The feature relates to a problem I've encountered as an active intermediate developer with managing the multidunious and varied queries that i might do both getting up to speed and solving problems in-the-wild. I find that what i learn and use doesn't stick right away, so rather than making the same query (and in this case sometimes getting different results) resorted to keeping 3 notebooks for each subject: a technical reference, a working notebook, and a learning log. That's a lot of notebooks!

I mention this because knowledge management and effective learning go hand in hand. And learning something you didn't know seems to be the problem domain of ai search for developers.

Organizing query results thematically by learning trees would be a gargantuan undertaking, and probably far outside the scope of what is already an excellent service. Just putting that out there.

Thanks!

Personally I hope it ends up either freemium or with a reasonable price ideally token based. I very much want to pay to incentivize development of good products but I already subscribe to enough things which is why I hope for more token based services.

(Of course if something is good enough and the pricing is reasonable there are sometimes room for exceptions. Kagi and Telegram are the latest ones for me and only Kagi is "safe", Telegram is mostly as a way to express my gratitude.)

One note, though: I’m finding the (mobile) UI very chaotic. The vertical spacing between blocks of text is irregular. And when submitting a follow-up question, it appears that nothing is happening - when, in fact, a new answer is being generated down below the links. I think revisiting some of the UI design could showcase the tech better.

https://phind.com/search?q=How+do+I+use+a+cache+that+is+like...

So I think these class of search engines should actually be called something other than generative AI search engines.

Maybe generative AI powered search summaries?

I asked it to write an essay on a subject I'm interested in, and an authority on. The essay, without citation, quoted whole sentences from both me, and wikipedia, verbatim. This seems like a bad thing. Certainly you can't use articles it writes if you don't know if it's plagiarizing.

And it all checked out. I think I’ll start using it more.

Reasoning, well my results were equal to google. Also your target audience are technical people, they can handle a more complicate rating process.

On the other hand, they did state it’s a developer focused search product for technical & factual questions. They’re aiming to make a bot that doesn’t provide opinions, or long convoluted conversations. On that basis your query isn’t a great representation of that.

But what should the answer be anyway? Winter? When there’s snow? When you have a break from work and enough money for a lift pass? When you’re feeling strong and healthy? Should the bot ask you clarifying questions to determine that?

If you asked a person this question they’d either ask you clarifying questions regarding what you actually want to know, or give you a vague answer based on where and when they like to go skiing.

Running this question again, I got:

> For example, ski resorts at Lake Tahoe usually open after Thanksgiving and close in late April, with February offering the best skiing conditions. In Colorado, the ski season generally runs from mid-November through mid-April, with February being the best time to ski due to the deepest base depth of snow on the mountains and plenty of powder still pouring in. However, early snowfall can bring early openings, and the weather can be unpredictable. In Park City, Utah, the best time to ski is from December to March, with January being the busiest month due to the Sundance Film Festival.

this currently returns a wall of text - 545 words for first paragraph. any way to chunk it up? or get a bullet point version?

I hate this so much, I have firefox nightly 107, how is everyone flagging me as out of date ... is it a bad library that you are all using?

I found someone complaining about a similar thing except on WhatsApp Web which also tells them to update Firefox even though they're on FF 107.

Maybe FF changed something used in common feature-detection checks. Perhaps it's fixed in FF 108+.

I think this is probably something that only cloudflare can fix

For everyone not on Firefox, this is one reason why we kept saying a browser monoculture is bad.

My intended answer is that you use the monitor as a ruler.

ChatGPT arrived at that answer after a few tries. This couldn't do it.

https://phind.com/search?q=show+ruby+code+snippet+to+search+...

Phind gets the meaning of computer vision completely wrong.

Google got the meaning right, but answers with a bunch of links of dubious quality.

Here is a random example: https://phind.com/search?q=how+to+list+BLE+devices+on+IOS+us...

I quickly got tired of asking for code examples in the follow up as it's an extra roundtrip.

Good luck

https://www.microsoft.com/en-us/bing/apis/pricing

Thats like 30-searches per day (including any tweaking). How can a solo developer make an MVP based on this?

Interestingly, the very existence of the new Bing does us a favor in this regard -- it warms people up to the idea of using something other than Google, even if it's just for a subset of searches.

We're thinking about ways to enhance readability by breaking the core of the answer into its own paragraph.

Great work

In many cases it would be nice if it had a TLDR at the top that's a single sentence, and all the details below.

Maybe it's just me, or maybe that what was being aimed for?