HNHacker News
TopNewBestAskShowJobs

estsauver

2,122 karma · joined February 18, 2013

Cofounder of ApolloAgriculture (https://apolloagriculture.com). We help smallholder farmers in subsaharan africa maximize their profits, and we're good at it.

Oregonian lost in Amsterdam.

earl at apolloagriculture dot com

[ my public key: https://keybase.io/estsauver; my proof: https://keybase.io/estsauver/sigs/2MlvHN12O0vg54og9F5h0xbq6eNMlONSt2O3AdFfV94 ]

submissionscomments
estsauver··on Show HN: In a single HTML file, an app to encourage my children to invest
So, bonds basically all tend to converge on the same risk adjusted yield. If you're seeing yields that look like this, the market believes the currency will slip or there's repayment risk (relative to USD bonds that are in the 4.75% range.)

Imagine you have a scenario where inflation is 0 in currency A and 10% in currency B. Would you rather have a 2% bond in currency A or a 9% bond in currency B? This is why Euro bonds go negative sometimes, when USD interest rates were very low and the Euro was deflationary relative to the dollar, it could push rates even further lower.

estsauver··on Cloudflare Email Service: private beta
So, interestingly, the places that actually did the worst in the Holocaust were generally the places where there were the least legal structures--even though you would expect it to be worst for Jews in Germany, it was often Poland and other states that had all legal structures and civil institutions destroyed who had it the worst.

https://www.amazon.nl/Black-Earth-Holocaust-History-Warning/...

estsauver··on Self-taught engineers often outperform (2024)
I came from a Chemical Engineering and Math degree, but ended up teaching myself programming while I was in university. I saw my freshman year roommate getting offered more for an internship than chemical engineers were making after they graduated and put 2 and 2 together.

I think there is a huge benefit to having had some sort of extremely rigorous quantitative background. When I talk with people who are self taught who haven't come from a hard (where that is an extremely hand wavey concept) can get started in engineering and often can be productive, but really struggle when things break or get difficult. There's a sharpening of the mind that comes with extreme analytic rigour, and I think that's more important than whether you get it in a real analysis class or in a CS class.

estsauver··on LLM Inevitabilism
I think it's worth saying that I basically completely disagree with your assessment (how you read the evidence, your conclusions, and quite possibly your worldview,) and think that if you were to give me access to infinite throughput claude code in 2018 that I could have literally ruled the world.

I'm not the most impressive person on hacker news by a wide margin, but I've built some cool things that were hard, and I think they are absolutely inevitable and frequently have the exact same "one shot" type experience where things just work. I would seriously reconsider whether it is something that you can't make work well for you, or something you don't want to work well.

estsauver··on Bitchat – A decentralized messaging app that works over Bluetooth mesh networks
FWIW, I've found a T-1000e to be a pretty good way to get people into meshtastic. It's not perfect because it has a weird dongle to charge, but it's pretty cool and I think you can convince people it's a worthwhile thing for emergency recovery.
estsauver··on Start your own Internet Resiliency Club
I think that's totally fair if you're primarily working in rural/not dense communities. If you're in a major city, in practice it just doesn't matter and ease of use is king. In a disaster, everyone with a meshtastic radio has it on. Your messages will propagate just fine.
estsauver··on Start your own Internet Resiliency Club
* I would consider adding a T-1000 device to the recommended list of devices, it's about the size of a credit card and works very well to add Meshtastic to phones. https://www.seeedstudio.com/SenseCAP-Card-Tracker-T1000-E-fo... It's a lot easier (in my opinion) to get people to stash a radio and remember how to power it to Bluetooth then to get setup from 0 on a new device. I think I paid about 40 euros each when I bought a pair.

* I have a Starlink mini--in the event that there is ever a broadly disconnecting event I'd be happy to share access to it. I keep it pretty much exclusively for emergency use and occasional camping/rural holiday house vacations. You might want to consider one too? They're ~250 euros new, which for someone who's starting a club for anything seems like a plausible expense. I believe there's a chinese version in case you don't want to trust the whims and emotions of Musk. * https://kiwix.org/en/applications/ is pretty useful if you'd like to have an archive of technical information, wikipedia, stack exchange etc.

* I try and keep whatever feels like the smartest open weight LLMs at the time available so if something real bad ever happened it'd still be available. I might add that idea to your preparedness list too--I'd probably take LM Studio with Gemma 3 over another random engineer on the Meshtastic channel :)

* Would you share channel config details for your IRC community? I'm happy to join.

estsauver··on Show HN: Eyesite - experimental website combining computer vision and web design
Hey, I just wanted to say this is one of the coolest demo's I've seen all year. This is really, really fun.
estsauver··on Build a minimal decorator with Ruby in 30 minutes
Doesn't this thrash the ruby VM's virtual method cache? (It's been ~7 years since I was working on production ruby code, but I remember it being a really painful performance issue for us when using dynamic method missing routing. I might be getting it confused with class extensions though.)
estsauver··on Launch HN: Vassar Robotics (YC X25) – $219 robot arm that learns new skills
For what it's worth, I wasn't able to pick an address outside of the United States when it still had units for sale. If you want a first international sale, I'm estsauver at gmail dot com and would be thrilled to be your first customer from The Netherlands.

Congratulations on a great launch!

estsauver··on Ask HN: What are you working on? (May 2025)
I shipped a tool that uses logprobs to generate pretty detailed estimates about how long tasks are going to take (i.e., give a new product manager/business person/sales person this tool and then they'll have reasonable estimates for how long various tasks will take.) If anyone wants to try it at universalestimator.com, it's free with code "TYHN".

I'm also making a picture book generator for my kids. I want to be from some pictures, a description of a story, and will then pop out a ready-to-print picture book for them.

estsauver··on Xiaomi MiMo Reasoning Model
Qwen3 and some of the smaller gemma's are pretty good and fast. I have a gist with my benchmark #'s here on my m4 pro max (with a whole ton of ram, but most small models will fit on a well spec'ed dev mac.)

https://gist.github.com/estsauver/a70c929398479f3166f3d69bce...

estsauver··on Qwen3: Think deeper, act faster
I don't think this is all that well documented anywhere. I've had this problem too and I don't think anyone has tried to record something like a decent benchmark of token inference/speed for a few different models. I'm going to start doing it while playing around with settings a bit. Here's some results on my (big!) M4 Mac Pro with Gemma 3, I'm still downloading Qwen3 but will update when it lands.

https://gist.github.com/estsauver/a70c929398479f3166f3d69bce...

Here's a video of the second config run I ran so you can see both all of the parameters as I have them configured and a qualitative experience.

https://screen.studio/share/4VUt6r1c

estsauver··on Qwen3: Think deeper, act faster
Is that Qwen 2.5 or Qwen 3? I don't see a qwen 3 on the aider benchmark here yet: https://aider.chat/docs/leaderboards/
estsauver··on Ask HN: What are you working on? (April 2025)
No, but it's on my roadmap! https://feedback.universalestimator.com/p/calibrationnorms-f...

Right now, it asks you some follow up questions and assumes you're a medium sized org, but I'd like to start to move that into configuration and do some sort of time/bayesian expiration of memory/information as part of the questions it asks.

I think a ton of the variance between teams is probably captured by some version of a few calibration questions, aka: - How large is the org? - What region is your org based in? - How long does it take to get the smallest possible change into production?

estsauver··on Ask HN: What are you working on? (April 2025)
universalestimator.com

I started experimenting and I think this builds out pretty neat estimations from jira tickets/other descriptions. When I was sitting in the CTO role, I spent a ton of time talking with people about how long/short various projects would be. When I was a developer, I hated the estimation piece because it felt like it was both keeping me from building and was almost never done with enough context to get really accurate results.

I was playing with the OpenAI API and I noticed that they can actually return a set of probability x next tokens and I thought that it might actually give you kind of cool ways to see the distribution of potential outcomes. (You can see an example here: https://universalestimator.com/estimates/c68db45b-7622-4bab-... that looks at a detailed ticket for implementing filtering on a dashboard.)

Let me know if you have any feedback, it's free with the promo code TYHN. If you run into any issues, please send me an email at earl@unbrandedsoftware.com

estsauver··on Prepper Disk
Part of the reason I bought a new macbook with a decent amount of ram was that I wanted to make sure I could have a couple of local LLMs available locally just in case for some reason the frontierlabs/other models eventually became restricted.
estsauver··on Show HN: I made my own TRMNL e-ink device
This is neat--what's the typical refresh rate/update speed for the image service?

Is the API is basically "Push PNG to server, PNG displays on whiteboard?"

estsauver··on Have we underestimated the total number of people on Earth?
This idea is extremely old, but also has a very bad record of being wrong for hundreds of years:

https://en.wikipedia.org/wiki/An_Essay_on_the_Principle_of_P...

It’s worth knowing if you’re proposing an idea that’s been continuously proposed, and has been continuously proven wrong by time and history, that you might want to consider if there are other drivers that continually oppose your thesis. I think technology is sufficiently powerful we’ll be able to power our needs for another hundred years at a minimum if we try.

estsauver··on “The closer to the train station, the worse the kebab” – a “study”
This "you have to choose D" ahead of time nonsense is why people distrust and dislike statisticians! Humans have priors on what is "close" that are independent of this particular article. If they had said "See, everything within 5000m" or "everything within 5m" you might have a point but "500m" being a rough definition of "close to a train station" is pretty reasonable.
estsauver··on Claude 3.7 Sonnet and Claude Code
The docs for Claude code don't seem to be up yet but are linked here: http://docs.anthropic.com/s/claude-code

I'm not sure if it's a broken link in the blog post or just hasn't been published yet.

estsauver··on OpenAI O3-Mini
I couldn't find in the documentation anything that describes the relative number of tokens that you get for low/medium/high. I'm curious if anyone can find that, I'd be curious to see how it plays out relative to DeepSeeks thinking sections.
estsauver··on Introducing Operator
I think it's just another way of accessing anything that doesn't have a traditional API. Most humans interact with things through the world with a web browser, with a keyboard and a mouse, and so even places that don't have any sort of API can be supported. You can still probably use things that define tool use explicitly, but I think this is kind of becoming a general purpose tool-use of last resort?
estsauver··on Operator research preview
I think one of the things I'm most excited for is that this really opens up, for practical purposes, a lot of websites that made it difficult to do things via API. For example, while I frequently end up booking AirBnB's, I find the process of searching for an AirBnB quite tedious.

I dream of a world where I can specify annoying things to me and build a perfect search for any house, that understands how I think about money, how I think about my family, and what I love and really extends how I interact with the world.

estsauver··on Official DeepSeek R1 Now on Ollama
Wrong post--This was meant for the anti-cheat post that's also on the frontpage.
estsauver··on TypeScript enums: use cases and alternatives
I like how this article demystifies TypeScript enums—especially around numeric vs. string values and all the weird runtime quirks. Personally, I mostly steer clear of numeric enums because of that dual key/value mapping, which can be as confusing as Scala’s old-school Enumeration type (where numeric IDs can shift if you reorder entries). In Scala, it’s often better to use sealed traits and case objects for exhaustiveness checks and more explicit naming—kind of like TS’s union-of-literal types.

If you just need a fixed set of constants, union types with never-based exhaustiveness checks feel simpler and more “ADT–style.” That approach avoids generating the extra JS code of enums and plays nicer with certain “strip-only” TypeScript setups. In other words, if you’ve ever regretted using Enumeration in Scala because pattern matching turned messy or IDs moved around, then you’ll probably want to keep TypeScript enums at arm’s length too—or at least stick to string enums for clarity.

estsauver··on Don't use Session – Round 2
Session’s decision to remove forward secrecy is another big eyebrow-raiser. Signal is widely praised for its double ratchet precisely because it limits the damage if a key is compromised. Session’s lack of PFS means that if your private key is ever exposed (or derived), all your past and future messages are fair game. For a messenger that aspires to anonymity or strong privacy guarantees, this is a major trade-off. I haven’t seen a convincing argument as to why the onion-routed approach can’t coexist with a forward-secure key ratchet.
estsauver··on Why is Cloudflare Pages' bandwidth unlimited?
Cloudflare generally seems to have a really smart strategy team. There's a really excellent Stratechery article about Cloudflare's strategy team more generally:

(Stratechery is down now, but the web archive is up.) https://web.archive.org/web/20250108182845/https://strateche...

https://stratechery.com/2021/cloudflares-disruption/

estsauver··on Show HN: Konty – A Balsamiq-alternative lo-fi wireframe tool for modern apps
The reason that I've heard used repeatedly is that a shocking percentage of folks who aren't Technology producers can't separate visual quality from "doneness" of a project. If you show some business folks something that looks like it works, they'll mentally update the project to "Nearly done!" and then everything else after that becomes "Unreasonable delays."
estsauver··on AI is coming for voice actors?
I don't know--I think we're going to see a rapid evolution here. AI's were previously not creative and many people said that was the hill it would never be able to climb, but it's hard to argue that Dali and Midjourney aren't art. I think literary & language forms are probably quite hard for it to master, particularly when it's not that sure if it will offend. (The content filter in aligning an LLMs response to "Not offensive" probably prevents it from ever generating jokes that would be in line with Ricky Gervais or Dave Chapelle.)

https://openai.com/research/gpt-4#:~:text=Visual%20inputs%3A... shows it at least explaining a joke.

← PreviousPage 2 of 14Next →