HNHacker News
TopNewBestAskShowJobs

james-revisoai

153 karma · joined April 16, 2020

submissionscomments
james-revisoai··on Ask HN: PG's 'Do Things That Don't Scale' manual examples?
I liked your story and technique in finding a need-now kind of audience with this method! And your bit about Square is underrated. Setting up auth/stripe/connecting everything can take so long you lose motivation. I like the idea of using Square.

I face a similar kind of situation in a similar market, would love to hear your advice on something, feel it may complement what you are doing as well - can be reached at lugg@revision.ai for this if you'd be up to chat.

james-revisoai··on Ask HN: SaaS Covers Too Much?
But why did people go from introducing their startups as "subscription" "freemium" or "B2C" to mostly all saying "SaaS"? I definitely feel this was not the case before 2013... there was a period when SaaS became terminology before it swept up these other terms... I feel it might be like machine learning/AI - ML may be a better phrase for the situation, but people just say AI now as it is trending...
james-revisoai··on Ask HN: Best no/low-code QA automation tool
It sounds crazy but combining two (e.g. cypress and selenium) can work fine especially with changed $browser config. So when I had this issue, I just did that for now, it's never needed more than 2 and been the simplest solution.

Cypress I think doesn't allow multi-tabs etc (or at least at the release version in 2020 when I switched to it, didn't, and I still mentally model it that way), so I go back to selenium for this kind of thing. It's not ideal to use selenium these days, but it works ok.

james-revisoai··on Ask HN: Would it help to have an extra senior developer for just an hour a day?
Three hours maybe, one hour is too few.

If I wanted a senior devs opinions, leaving them days to ponder and consider their responses is also valuable, so three hours every few days is much more valuable than one hour a day.

The problem with one hour a day is like having a washing machine once a month: It's useful, but you still need to rely on the workhouse alternative for 3/4 weeks, and it saves you time, but might add more mental burden or risk, making it risky to do.

james-revisoai··on Ask HN: Extracting Knowledge Graphs from LLMs
Look at the Kaggle competitions that use embeddings, such as those by the Learning Agency. There was also a paper two weeks back and one two days back about inverting embeddings back to original semantic forms.

LLMs are very good at dealing with contextual polysemy, the catch is that an embedding of say a topic, will be far from the topic in it's different possible context. So a knowledge graph would be possible, but how you would find these possible areas, or why you would constrain it, is sort of another question.

Now if you are just asking about education, you can get it to generate lists of relations, and map those in concept maps etc(quite a few tools do this), but that's pretty superficial, as such...

FWIW back when LLMs were more prone to hallucination, the convex hull of the embddings of known ground truth statements was more likely to contain truthful generations and relevant generations than those outside of it when I worked on a quiz-generating application in 2020/21 doing this.

In my opinion though you should try to embrace this malleable nature rather than constrain it...

james-revisoai··on Ask HN: How do I start over at 37?
Show genuine interest in a project, be better at meeting deadlines (by overestimating) than others, show social skills most in tech are weaker at.

There are many edges you can have compared to most tech folk. Why not try using ChatGPT to begin some python and html coding and make some things you can get excited about when showing people?

You don't need social capital for a more entry type role, so don't overthink this

james-revisoai··on AI hype is built on flawed test scores
Yeah you have hit the nail on the head here. A lot was predictable with seeing that GPT-2 could reasonably stay within language and generate early coherent structures, that coming at the same time as instructions with the T5 stuff and the widespread use of embeddings from BERT told us this direction was likely, it's just for many people this came to awareness in 2021/22 rather than the 2018-2020 ramp up the field/hobbyists experienced.
james-revisoai··on AI hype is built on flawed test scores
A lot of the progress in the last 3-4 years was predictable from GPT-2 and especially GPT-3 onwards - combining instruction following and reinforcement learning with scaling GPT. With research being more closed, this isn't so true anymore. The mp3 case was predictable in 2020 - some early twitter GIFs showed vaguely similar stuff. Can you predict what will happen in 2026/7 though, with multimodal tech?

I simply don't see it a being the same today. The obvious element of scaling or techniques that imply a useful overlap isn't there. Whereas before researchers brought together excellent and groundbreaking performance on different benchmarks and areas together as they worked on GPT-3, since 2020, except instruction following, less has been predictable.

Multi modal could change everything (things like the ScienceQA paper suggest so), but also, it might not shift benchmarks. It's just not so clear that the future is as predictable or will be faster than the last few years. I do have my own beliefs similar to Yann Lecun about what architecture or rather infrastructure makes most sense intuitively going forward, and there's not really the openness we used to have from top labs to know if they are going these ways, or not. So you are absolutely right that it's hard to imagine where we will be in 20 years, but in a strange way, because it is much less clear than in 2020 where we will be in 3 years time onwards, I would say it is much less guaranteed progress than it is felt by many...

james-revisoai··on Ask HN: What's the best way to add search to my website?
This thread makes me feel old... I remember seeing "Google Custom Search" with Custom Search in grey on so many websites just maybe 12-15 years ago - for it to be so unusual that people on HN don't tend to know of it, seems so odd...
james-revisoai··on Ask HN: What are you optimistic about?
Close agree on 1 and 2, not well educated enough on 3.

Current mechanisms of computing and the way things like security build up from them seem so unnatural, so unbeautiful. I think we'll see something soon.

james-revisoai··on Tell HN: Postman update removes all your stuff if you refuse to create account
This happened to me 2 years ago, I think it has always been a bit of an issue.
james-revisoai··on Ask HN: Will Bootstrapping See a Resurgence?
I suspect your best bet is to reframe your solution if the technology or market need can present a gap on a chart, as a whole new startup

e.g.

"b20000s app/standanrd $industry app": 60% success b40000: Shiny and Fancy. With AIMLGPT machines. And all. 80% success"

Basically you are probably lacking urgency and humans tend to rest a lot of assumptions on some urgency, especially for business.

At least don't start by introducing the existing business. At best say you've spent x years as a CEO/CTO etc, and pull your experience in to answer questions, as well as clarify the situation if needed.

james-revisoai··on Ask HN: What's the best local sentence transformer?
Like caprock says, e5 are the best tradeoff for model size/speed of embedding for the results you get, they will be great at semantic search etc in English.

Possibly consider cross-encoding for semantic search depending on your use case, but whenever cross-encoding is useful, generative embeddings like Ada are usually much better... There used to be embeddings useful for things like classifying sentences are entailing one another, whether a sentence was complete, but these are basically completely supplanted these days.

Do consider the all-mini type embeddings (default in sentence transformer) for speed or on-device use. They are half the size (and therefore less than half the computing for distance functions) so they are faster for large searches etc, which is useful if you run your own stuff with vector stores rather than a service.

james-revisoai··on Need advise on job hunting after a long career gap
Good news is, it's a time of change and goldrush, and salaries have significantly increased since 8 years ago.

Bad news is, what you have learned seems to represent what is now in declining state, the technology that was more useful before ChatGPT (this may be subjective). Python, sci-kit learn, numpy have been made more and more obsolete by ChatGPT itself taking over many custom model roles (that sci-kit learn, pytorch or spacy would have helped you do), and chatgpt can simply code stats tasks/dashboards so rapidly there is much less employment focused on this than a few years ago. Lots of data engineering is useful, but for analytics rather than predictive (machine learning tasks), since the predictive stuff, you'll probably find you don't need things like feature detection, normalisation etc for depending on how willing you are to try the cutting edge tools coming out.

Why don't you considering combining the two at the twine to your advantage - Have a look at ChatGPT and using add ons or other recent technology (e.g. Code Interpreter) - you could be in the prime and ideal position to take advantage of this shift, without preconceptions about exactly what should be the "next generation app".

james-revisoai··on Just kidding: the evolutionary roots of playful teasing
I don't think cats only do aggressive teasing (presumably, you are focusing on how they can treat other cats or rats/mice)

Cats can be found playing with weaker animals like ducks or rabbits, and comfortable yet playful in the same nature as puppies with dogs and puppies. Lots of this teasing isn't of agressive nature, but it usually happens within the home/safe place for the animal

james-revisoai··on Ask HN: Lessons learned from implementing user-facing analytics / dashboards?
Use colours and graphical elements (generated graphs), but:

Obey rules of spacing more carefully than other rules to avoid overwhelming.

Do not use colours unless signalling information, so users can be alert and relaxed when needed.

As soon as you have more than 2 types of information, have expanding panels, which remember whether the user expanded/collapsed them.

Lastly, remember that speed of loading data is much more important for dashboards in general than a random page. Cache data, or initially load only summary data, or only load the latest day by default and then fetch the weeks data. Remember clients may make purchasing decisions based on how fast your stats page of your SaaS usage loads when they are showcasing it to their C-suite, and a 15 second wait can cost you your enterprise sale.

james-revisoai··on Ask HN: How do you use AI to get things done faster?
Its helpful to do boilerplate in languages or legacy code you touch less often, e.g. React class components.

But also, it can reasonably set types for the inputs of say a React component, just based on an example (made up) usage of some props you give it. And you can even, if you are very uncouth, simply generate the sensible props for a certain kind of object given a description by say the user.

It can do things like generate controlled form state management etc, which was quite error-prone to do with macros when things get complicated.

It can summarize commits for documented code accurately too, especially if you comment what you've done for some reason or another. This saves a good 2-3 minutes an hour I didn't know I could save, for the average of say 1-2 commits I'd do usually per hour.

Notably GPT-4 is needed to do these consistently well.

james-revisoai··on Ask HN: Why aren't there Open source embedding models with context length > 512?
There are as mentioned, but additionally, for many models, you can split content up into several vectors (say one for each sentence or paragraph depending on how the model is trained) and pool the vectors together to get a representation that will span the content overall well.

Since the models trained to work on single sentences (like Mini-V2, the SBERT default) work worse at length, pooling representations of sentences is typically more useful.

For deliberately longer representations, generative model embeddings or document embeddings are the right answer sometimes.

james-revisoai··on Ask HN: Why did Python win?
Ruby was more popular, especially when Rails exploded. It was great for scrappy startups, that didn't need Java level quality (and costly time) nor the risks of PHP 4.

The python modules eco-system lead to the stake change. Scikit-learn, Pandas and Numpy made it so useful for visualisation, research, old school modelling ML. Writing and saving files was simple. Making your own module took 2 files with 4 lines of code. How easy.

After establishing use cases in Python 2, which already allowed a lot of C code in practise for speed, Python 3 changed everything with a simpler language that makes it clear what is normal and weird. You might dislike [:-1] at first, but you definitely know it's weird compared to other languages, so you can know what to search for. Not true of abstractions in some other languages like Ruby, for migrating devs.

Then Python caught with Tensorflow/PyTorch a real wave of being the place to go. Kaggle involved 100,000s all coding in Python for data science in the latter half of the 10s. It did this at the same time as the LAMP stack and LAMP pre-installed VPS running cPanel were dying to React/node.js and PaaS approaches on the one hand, and big cloud providers on the other. So suddenly, there was not an easy default 'nix environment with PHP to rely on being available so easily, and adding python to an existing server is super easy (and Flask/Django easy to grep), making adding a backend for ML or whatever, very easy too. Ruby wasn't even on the playing field for many of these use cases.

I think a last small impact was that Ruby was misused where shell scripts could have worked. Many people made Ruby cli tools for their internal team, when shell scripts would have been fine - this frustrated some of their peers away from Ruby. There is no difference from Python here, but I just noticed it more.

james-revisoai··on Earning the privilege to work on unoriginal problems
Any more links to this / discussion on this?

Would love some talks covering where the adjacent areas are to the cloud offerings - are you thinking at AI inference startups or more dev tooling?

The oddest trend I have noticed (may be selection bias) is how many people build dev tools these days... to the point even non-devs are starting to talk about their no-code builder startups... it's getting crazy.

james-revisoai··on Ask HN: What does your self hosted stack look like?
Good advice. You threw me back into the memory bank with this!
james-revisoai··on Ask HN: How do you look for jobs in 2023?
You could always try leaving the FAANG name off or adapting it to the project name, and see how things go, given the associations people have with FAANG outside of SV/HN nowadays. Helped my friend a lot.
james-revisoai··on Ask HN: I learned useless skill of prompt engineering, how relevant will it be?
About as valuable as autoregressive engineers who wrote code like 'const summaries = ["' into prompts in the days before instruct/RLHF based models like ChatGPT, a skill now not needed.

That it to say, in medium time, no meaningful benefit.

james-revisoai··on Ask HN: Burnout because of ChatGPT?
"new developments and new tools and frameworks and models keeps me feeling like I am stuck in quicksand with lead boots while all these developments are whizzing past me"

Exactly, even as somebody with a background in NLP, some things that I developed over the course of two-three months in Summer 2022 were promptly replicated and nicely put into accessible, open source modules by winter that anybody could run in a few lines of code, besides some bespoke bits like processing (QA on documents, etc). After a few such experiences, you begin to become wary of working on the edge, since the progress is so fast you could waste lots of time on convergent evolution. It really is hard to predict any part that won't succumb to that in the current racing stage.

james-revisoai··on Ask HN: Why is there no API for testers on demand?
QA at a software company is a bit far from what I'm thinking of here - I'm not looking at finesse or QA for a product facing a large audience; more security against regressions.

Think small side projects, one off mom-n-pop sites, shopifies or wordpress or drupal with extensions or plugins that can clash.

Fiverr and Upwork both deliver too unpredictably, can take anywhere from 30 minutes to a day. If you have 5 changes to get tested it's a nightmare doing it with those services. With scale you can hire a tester but if your work is come-and-go it's not sensible and sometimes you want fresh eyes on each new version.

With tools like vercel I am surprised there's no way to programmatically include paying a tester - like a CI - to give you the human-touch manual all clear, without having to hire somebody.

james-revisoai··on My frugal indie dev startup stack (2022)
OP, do you have one or two examples of Ahrefs being used in practise for a startup?
james-revisoai··on Ask HN: Why is there no API for testers on demand?
Problem is they don't reliably get back to you within a short time frame, sometimes you have a deadline or multiple features and just want that reliability of getting a response same day you can get with an internal test team, you know.

hhm. I feel like vercel etc will make such things easier in the future. it feels like a micro version of the uber-taxi unreliability problem

james-revisoai··on Ask HN: Why is there no API for testers on demand?
(and to anticipate the answer of Amazon MTurk - I don't think you can easily do it task specifically to each individual or start a task through an API/email?)
james-revisoai··on Ask HN: What are the capabilities of consumer grade hardware to work with LLMs?
Well, if you looked at the AI world in 2019, the models were mostly less RAM intense (typically 0.8GB-2GB RAM for BERT/T5 non-large for example). And computer vision models always hit the max you'd throw at them.

Things twist and turn. Text requires a large amount of parameters to be useful. If we as humans required 8000x8000 pictures to discern objects, images would require more. It's quite anthropomorphic. That's really the core of it.

james-revisoai··on Ask HN: Share one man boring internet businesses
I've seen you comment before on related threads; congrats

The question that always comes to mind is, how do you get distribution and reach the market?

There seem to be so many competitors.

← PreviousPage 2 of 6Next →