talolard

Simple can mean a few different things: fewest moving parts, easiest to code, minimal amount of stuff to learn… I’m good with Django. You can implement the app in Django and use Celery to do the scheduled work. I think that’s easiest to code

Someone on the internet is wrong!!!

Since onsite is usually implied and rarely mentioned explicitly in the text, taking a corpus that mentions either term will bias heavily towards remote

I subscribe to an alternative premise - "The browser is effectively an OS and thus a standardized distribution target". There are cases where the people I want to distribute to don't suffer from the constraints you mentioned. e.g. will tolerate a 20 second page load in exchange for the experience my product will provide.

Being able to reach all devices and have a super fast experience without consuming bandwidth is always good, but not always a high priority .

Worked at a large bank and saw /worked on a few interesting projects.

Obviously a lot of work on trading. One group of market makers had an RL agent built that would trade small sizes. The value here wasn't in profitability, it was that it freed up the traders to serve larger tickets while being catering to a broader swathe of the market.

Another group dealt with underwriting commercial loans. In that process, borrowers submit hundreds of documents that need to be classified (This is an architects license, this is a pest inspection report ... ). The data was too varied for simple heuristics, but fairly straightforward NLP eliminated a good chunk of work.

If you extrapolate, a lot of the problems I've seen "the average" (not Google) company solve with AI is optimizing an internal process, as opposed to making a new product offering. So "huge success" these were not, but they were successfull. I think that for the average company, expecting "AI" to make a "huge success" in terms of business impact is often a sign of weak product thinking.

Hi HN, I wrote this and wanted to share a few more words. This post discusses a trick for training neural networks. Usually when we train a NN, we just show it what the correct answer is, but we don't say anything explicit about the wrong answers. This trick, Complement Objective Training, is a way to give the model "feedback about its mistakes". In practice, it makes the model better and train faster.

The important thing is that it's pretty easy to implement. Many cool research ideas are a months work, but this takes 15 minutes with pytorch lighting so it's in the realm of low hanging fruit.

Isn’t the middle ground writing custom allocators ? Allocators allocate blocks continuously and developers keep developing with standard OOP?

I made a dev tool into a small business. Came hear to rant: invest in ui upfront. At the end, customers visit your site and you have very little time to make them go wow. Your spending a lot of attention capital when you don’t have much

I'll offer an illustration. We sell our SaaS at X. We have a customer that needed stronger SLAs and servers located in Europe. They pay 8X and (I think) both sides feel like they got a great deal.

So to paraphrase Arubis, this is a business question,not a tech question. At what price point is the technical investment covered ?

Congrats! I frequently deploy sklearn models to lambda with API gateway so this is spot on for me. I hate integrating lambda with gateway, building custom layers etc so this scratches a real itch.

I’m not a fan of deploying my models on your infra, but maybe that’s an opportunity for you. I’d pay a subscription if the tool would let me deploy to my aws account. I think you can sell a cloud formation template through the aws marketplace as a manageable way to execute that.

Anyway, looks awesome and definitely scratches a real itch. Good luck

While a great effort, the author conveniently forgot to mention their commercial competitors and even some open source ones. I’m from LightTag, we should be there as should explosion’s Prodi.gy and Labelbox. And where are docanno, Webanno and Brat?

If this were a random community member I wouldn’t make a stink but Heartex is a commercial company with an open source GTM.

This is a profound insight.

Dataish person here. If you search for Ml/ds you’ll get technical content that won’t help you much or fluff. You’d be better served by looking at companies you know do DS and searching for business problems they solved with it. Examples: Amazon boosts sales by recommending products. Visa stops fraud by doing fraud detection. Walmart manages inventory by doing sales forecasting .

Once you’ve found a use case that could be relevant to your stakeholders you can dive into the technicals

As a data scientist who does more frontend, I think this is a really valuable concept. Hello by users/stakeholders engage with our work is the way to push it forward in the org and a dash of frontend can do wonders for getting that message across. It’s wonderful that people are making resources about the frontend for data scientists

As a small vendor, I hate procurement platforms. Vendor UX always seems to be an after thought, since you’re the pipe between our customer and getting paid, why care about us. I hope you guys do it differently, maybe be so good to vendors that we champion you to your prospective clients. I hope you put bill.com and buy2pay out of business

I have a question on the compute aspect regarding your business model, hope I’m not being to nosy..

I tried HL, the experience was stellar (well done!) and it made me think...

To get AL working with a great user experience you need quite a bit of compute. How are you thinking about your margins, e.g the cost to produce what you’re offering versus what customers will pay for it?

I wrote a blog post a few years ago about possible downsides to active learning.

https://www.lighttag.io/blog/active-learning-optimization-is...

To be fair to the Humanloop folks, the criticisms probably don’t apply to the kinds of models their using (modern transformers).

I have a few suggestions regarding search performance.

As others have mentioned, there isn't one "right answer" but there are a few nifty tricks you can use to boost performance.

You could try a trigram index[0] as the primary search index. You'd loose stemming, but gain language support and indexed regular expression search which is a nice power user feature.

Looking at your where clause, did you index language and textsearchable_index_col together ? If not, you are giving the planner the opportunity to skip your full text index.

if you order by ts_rank_cd then the query will sort all rows, including those that don't match. A better pattern is to take the ranking score as a column and then sort it in a subquery. [1] from stack overlfow has an example. (As an aside, from pg2, CTEs are no longer an optimization fence, so you can write the query out more cleanly with a CTE and still get the desired performance).

You should experiment with GIN vs GIST indices. GIN are larger, so on a large dataset the index might not fit in memory. You could pay more to have more memory, but worth trying a GIST index to see if that makes things faster just because it fits in memory.

A final frontend comment, I'm a fan of infinite scroll for this kind of stuff. You already have pagination effectively set up, you could spoil me as a user with infinite scroll. react-virtualized[2] is a great library for that.

Hope that helps.

[0] https://www.postgresql.org/docs/9.6/pgtrgm.html [1] https://dba.stackexchange.com/questions/4054/ts-rank-massive... [2] https://github.com/bvaughn/react-virtualized

Question for OP Can you share how you manage the borders between the open source product and the infrastructure for offering a SaaS

As a trivial example, I’m assuming youre using plausible in plausible, do you have protections ensuring secrets don’t leak to the open source for history ?

Yeah, Gatsby-image was what drew us into the platform to begin with. That's what made it fast to get started.

I found the plugin system confusing, and poorly documented so it was hard to know how to get the most out of a plugin and adjust it to whatever our own funny use case was.

The graphql thing was annoying, we don't use graphql and don't know it and didn't find any advantage in having it around.

It's been a while since I touched it, but every time I need to a put it off because it's not a nice env to work in compared to our "from scratch" react apps

Poor Gatsby. We run a Gatsby site, it was great to setup fast fast fast and has been a pain in the butt ever since.

Next, from our limited experimentation, has much better ergonomics and it looks like the company has a stronger grasp on go to market.

Crazy how fast that happened

If you have schools (e.g. institutions, not individuals) already paying, maybe experiment if you can make a pricing tier for institutions. E.g. Special School package 3k/yr includes unlimited tracks, group by classroom, some other school specific feature ...

Hi, Congrats on hitting that milestone . As others have said, the site looks great and it's clear you've put a lot of time into it. I can offer a few feedback points:

The pricing tiers could have better names, from a quick glance I wasn't sure which plan fits me. Maybe take a look at what are common traits for customers in each plan and see if you can come up with a one word persona for each plan that way.

The product tours after signup are really clear and crisp well done. I've found that getting users to Wow fast is key to drive conversions, and that tours aren't enough. They show me how to get to wow but don't bring me there.

A common pattern is to prepopulate the app with some data and lead the user to a likely wow event quickly. That's super hard, because you need to know what wow is, but it's gold when you crack it.

I've asked customers "What made you decide to give us money" and that can get some good insights towards what wow is.

I just visited the site for the second time, and my dashboard is empty because I haven't done anything. A real user will likely churn, because they don't know what to do. Nudge them hard towards the action that will bring them to wow.

I'm parroting patio11, but your pro plan is priced to low. Try raising prices on it (a lot) and see if anyone goes for it. Anecdote: I set a pricing page to have $50, $180 and $1200. $50 sold the most, but one day someone self served themselves an annual $1200/month plan, which is 24 $50 customers. Point being, if someone wants to give you a lot of money, give them the chance to do so.

Last point, I just noticed the video under the the pricing section. My personal opinion is that it's weak for marketing purposes, it tells me how to use the app, but doesn't immediately show me what the app can do for me. The point of a video on a marketing page is to get me excited to try and then convert. Make me feel excited ... On the other hand, it is a very good tutorial video, maybe put it on the dashboard of the app if the user hasn't uploaded any tracks yet.

Good luck!

If I understand the paper (which is questionable), that's what the author is aiming for. E.g. he's saying 1) We can make these amazing black boxes 2) We don't really understand them 3) But when we make them with gradient descent they end up being almost kernel machines 4) We know a lot about kernel machines, so we can use that to "remove some uncertainty"

Hmmm, I think that's not precise and my use of "architecture" was misleading.

If we're thinking in terms of "universal aproximators", an RNN is a way to make a sequence of approximate functions for a sequence of inputs.

But it's still a sequence of functions, not a single function.

For a 1 layer network to have the same ability as an RNN (take an unbounded amount of context) it would need to have infinite width which is a no-go.

50/50 I can shed some light or am way over my head:

The point of the paper is that if you train a deep learning model with gradient descent then the resulting model is effectively a kernel machine, regardless of model architecture.

The nice thing about a kernel machine is that it is simple (just one hidden layer) and we are able to use to analyze a kernel machine more effectively and conveniently.

So, I think the contribution here isn't "these sets of universal aproximators are equivalent" but rather "We have this effective but opauge deep learning thing, turns it it's actual a kernel machine in retropsect so we can bring 'kernel tooling' to analyze the deep learning mode"

There’s a Reddit dump on big query so you can do this idea at relative scale with sum sql

I really love postgrest! I did a lot of django before and found myself constantly wrestling with the orm to express what I wanted, then writing endless pages of serializers and view sets and had to debug why they were always slow.

Postgrest has a learning curve, but the performance boost vs django is huge, and I can use more advanced db features like triggers more easily and in a way that’s native to the overall project.

This described is perfectly

Buried in there is that we finally get named capture groups on regular expressions. What a day!

Hi I started coding at 27 and am doing well despite the late start. If you like coding keep investing in growing your skills and try to get a job/internship at a place that aligns with your interest.

If you enjoy it it’s a great job with great money, and you get to learn about whatever domain you’re coding for. IMO it’s an implementation skill and having a solid academic background in biology would still make you a more effective biologist