It told me my ~10 year old js project was 50% AI generated. Yeah, this is more or less the same as "AI text detector" stuff that won't work reliably (but people who don't understand LLMs will still use it to blame others)
Built vaporlens.app in my free time using LLMs (specifically gemini, first 2.0-flash, recently moved to 2.5-flash).
It processes Steam game reviews and provides one page summary of what people thing about the game. Have been gradually improving it and adding some features from community feedback. Has been good fun.
I usually find that if a game is rated overwhelmingly positive, I'm gonna like it. The moment it's just mostly positive, it doesn't stay as a favorite for me.
Those games are usually brilliant - but those are very rare. Like "once in a few years" kind of rare IMO. While that is a valid approach, I play way more than that haha!
What I found interesting with Vaporlens is that it surfaces things that people think about the game - and if you find games where you like all the positives and don't mind largest negatives (because those are very often very subjective) - you're in a for a pretty good time.
It's also quite amusing to me that using fairly basic vector similarity on points text resulted in a pretty decent "similar games" section :D
That rating is not (just) a function of positive to negative ratio. Small number of reviews (ie small games) can't reach that rating although they might be equally well received.
I've started self-hosting mailcow (dockerized) on a Hetzner VM ~6 years ago or so. It took about 6 month to clean the rep of my IP address as it unfortunately been in spam lists previously - but after that, it's been a pretty smooth sailing. And all for a measly ~10 euro/mo including daily backups.
To finally figure out how to effectively design and test MVPs / prototypes without overdoing it. I think this has been a bane of my startup life for pretty much whole ~14 years I've been doing it.
As someone who’s on research, I do hundreds of prototypes each year. My time optimizations include having a backing hypothesis for each prototype, and prototyping one hypothesis at a time, even if two hypothesis touch on the same domain problem, I spin a fresh new prototype for each. Usually I’m able to validate the hypothesis even before finishing the prototype, as merely getting the hands on the process reveals the answer. Also, each hypothesis lend itself to some specific kind of prototype. Some things will be amenable to code prototypes, others to statecharts, others to visual designs, etc. so getting intimate with your toolset is essential for productivity, as most time is spent with these tools. Even simple things like mastering keyboard shortcuts represent significant productivity gains. But building is not the hardest part, by far. The hardest part, for which there’s hardly any optimization is this: how to filter and select an hypothesis to begin with? Overall, I would recommend getting good at prototyping in general (as a mindset thing) and not simply trying to quickly make a particular prototype.
Having a research background - for me, doing this in research was always easier. Because you generally have very straightforward ways to validate your hypotheses.
Validating things with customers - in my experience - can be extremely tricky as they might not even know what they want
Absolutely. In this case this means your pre-process will be less deterministic, hence your success ratio will be lower (maybe a lot lower). Perhaps improving your communication and data analysis skills would do good, as this sort of context is characterized by asymmetric information. Even though you’re in the market, this is still research, just of a different kind. I would try and find the “hidden variables” for each candidate hypothesis. Assuming “performance” to be the true abstract metric you’ll be striving for, it’s important to identify what particular concrete performance metric is worthwhile improving. For instance, consumers may suggest interest in feature A and B, but once you analyze their yield you identify they are time-saving features at their core. If time is the hidden variable, then where in the product lies the best opportunity for saving user time, even though the customers might not be mentioning it? I would guess that most requests or indications coming from customers will fall more or less in the same few buckets in respect to hidden variables and related metrics. Overall you shouldn’t take customer feedback as prescriptions, but as symptoms, and those symptoms might not even be related to what they’re being vocal about, since they might be stuck in a corner not by the lack of improvements in that existing domain but maybe by the lack of an entire novel domain. Think of it like a Tetris game, locally, each region in the bricks stack require some particular piece to solve that region, but globally, that’s not a good strategy, as a whole different piece in a whole different region would solve a larger portion of the stack.
Then there’s, of course, the issue that the whole of current customers feedbacks might be blinding you from what the product really needs to grow beyond that base, since a startup’s ultimate goal is growth, and fulfilling the current customer’s needs might make them happier but result in zero overall growth.
+1. I'm in the same boat. Making a conscious effort to build quick and dirty prototypes when needed and only build them after I get a potential customer to jump into Figma with me and spend an hour of their time designing something with me. If they're willing to do that the hypothesis is they actually really do want it.
We're using top level wikidata classes (you can see specific classes in JSON response). Full list is not published yet, but will be available in the near future.
Edit: to be clear i'm not buying this for a second. Your 1st demo produces the exact same output as Babelscape/wikineural-multilingual-ner demo. If you built off of their model you need to abide by their license of non-commercial use and share-alike. I think you're a bad actor and I can't believe HN is leaving this up. Your company name also has negative implications: https://en.wikipedia.org/wiki/Borg
So, er, how exactly would I prove it? Do you want me to share our code / db / etc we've worked on for past ~two years? :)
For one - how many wikidata classes exactly do you get from Wikineural? If I remember correctly, it can do four (person, location, organization, other). Our models do several thousands.
It'll likely annotate similar things in text since our model is also transformers-based (which is basically current state of art) - can't really do anything about that.
Yes? I think my stance is very clear. It appears that you have forked some opensource models and built proprietary software off of them which you are now trying to profit, against their licenses.
