Dug into the available information about Manus's business. I imagine this is pretty reflective of the new class of $100m ARR AI companies. High CAC, high costs, high competition and no real clue where things will land long term.
Routes you to a random LLM - just a fun way to force yourself to try new things not just rely on muscle memory. Currently only has 5-10 and roughly weights the routing by popularity (ChatGPT -> Gemini -> Claude .... Copilot). Have been using it personally but thought I'd share in case folks had other ones to add. Main learning so far is that many are really good. Often small UX decisions can lead to better experiences than just AI horsepower.
I believe it's not strictly more decks, but also how often they shuffle. If they add 7 decks but don't shuffle until near the bottom, it's actually advantageous as the count can get very skewed (e.g. 50 cards left and 30 facecards remaining). However, with 7 decks and reshuffling halfway through the count doesn't reach as much relevance to the next card.
I wonder if in an application you could branch on something more abstract than tokens. While there might by 50k token branches and 1k of reasonable likelihood, those actually probably cluster into a few themes you could branch off of. For example “he ordered a …” [burger, hot dog, sandwich: food] or [coke, coffee, water: drinks] or [tennis racket, bowling ball, etc: goods].
My amateur understanding is that the benefits are:
A) H2 is exceptionally difficult to store at scale. Salt caverns are an option, but that is limited
B) methanol can be used off the shelf by existing generators so the prices are much more economical and they can use both fuels during a transitory period
I mean, it's totally fine for a non-profit to "wallow in irrelevance"? I think the fundamental issue here is that a tax-exempt non-profit organization like these board members are leading _shouldn't_ necessarily be chasing fame/fortune. By definition they've put mission above profits (with tax writeoffs as a benefit). The naivety seems to be that the company is trying to be altruistically mission driven while also acting like a typical ambitious/profit seeking startup. Startups are great! Just weird to wrap one in a non-profit which brings different incentives
Between open source modeling tools being incredible, transfer learning allowing dirt cheap fine-tuning and now mega-models being able to instantly give you a "mostly right" data set, the cost of creating ML features has dropped to almost nothing.
Products that took quarters/years and required big budgets for labeling, ML specialist, GPUs etc just a few years ago can now be done in an hour or so for free (if you are scrappy). I imagine this is going to lead to a ton of great ML features that weren't worth funding in the past but are very valuable in aggregate. Similar to the mid-2000s when the cost/ease of web development came down enough that there was a lot more experimentation and fun to be had.
Clever idea. I think you would have to recompute the context (ie embed the prior tokens) every time you swapped models because the weight distributions would be different for each model. Going from big->small might make this overhead worth it, but going back from small->big would assuredly be very costly.
Cool stuff! But it’s criminal to not call attention to JQ’s elder sibling CSVKit. It’s invaluable for playing with csvs. Much easier to parse out columns, allows you to generate new csvs and even merge them. More importantly, it allows SQL on csvs (via SQLite iirc) which empowers all sorts of csv shenanigans. The bash scripting this enables us incredible (good and bad).
That video did the same for me. I also like the reducible video for Fast Fourier Transforms as well as the Veratasium piece that shares some fun history.
I too have always been blown away by Shazam and pondered how it could possible index so much content for fast lookups. A few years ago this article was super helpful in helping me understand and learn a lot. Fun read which required a lot of side googling for me
http://coding-geek.com/how-shazam-works/
The sequel “Dark Sun” is also quite good. Continues the story into the making of the hyrogen bomb, the Teller-Ulam design, and the USSR’s efforts. Much shorter too :)
Yes, with SharesPost, equityzen and Forge. Company had a pretty standard policy so approval was straightforward but took a few months and required some legal opinions (costing $,$$$). In my case the deal was for the shares directly. Biggest lesson was really just that the fees really added up (near 10% if under $100k) and it’s a shame more companies don’t do more to broker these deals on employees behalf to get the best price and cut out the middlemen (another employer did this and it was great).
DM if I can help with more info, experience with each company was essentially the same and you should just reach out to all of them to see who has inbound buy requests at the highest price. Worth noting that there is more room for negotiation on price than you might expect.