For sure true for specialized ones like MedGemma (healthcare). In my testing, the 27b model is at least on the same level as frontier, and in some cases outperforms them. 4B is insanely good too for some lighter workloads. Thanks G for working on this!
Congrats Andrej. Let me know if you are looking for someone to take over Eureka Labs. We were in the same WoW classic guild. domrdy on x. We can duel for it :). Also, I need $10M if any VCs are reading this.
I agree with you. I’ve failed many live-coding interviews because I start focusing on how the interviewer perceives me instead of the task at hand haha. Hiring managers still want to hear the candidate talk through a technical problem. I built a way for candidates to screen-record their take-home solution using only their microphone and screen share, with as many takes as they want. This lets hiring managers hear how candidates communicate and explain their technical implementations, a skill that matters on teams. But it means managers have to watch those videos. Many of our customers use it and candidates like it, but it still takes extra effort on the candidate’s and hiring managers part, with no guarantee that it will "pay off".
I think certificates are interesting in theory, and there’s an entire industry built around them. AWS does a solid job with its certifications: if you earned one five years ago, it’s still more or less relevant. I'm not sure how certifications would work for proprietary technology.
Labor laws in many places are less “flexible” than those in the US, and they don’t support what you’re proposing, for good reason. I wouldn’t quit my job or uproot my family just to convince a manager I’m worth keeping.
As with everything, it depends. Live coding interviews work. They’re not the best candidate experience, but they work at Meta, Google scale, minimizing false positives better than most other formats. What makes them stressful is the lack of interviewer training and the abstract, puzzle-like nature of the problems, which you can really only solve if you’ve spent time studying (e.g., LeetCode) or you’re fresh out of college or academia.
I’ve worked in the assessment space for 6 years and have seen many hiring processes, from Fortune 10 companies to startups hiring their first engineers. The range of signal required, "how much time can my engineering team spend with a candidate", and how much candidate experience you can get away with, is huge. I’ve also been a candidate myself and failed many live coding interviews. It made me feel terrible about myself. The last time for a role at Ycombinator (the interviewer was super nice).
When I work on my product, I try to view it through the lens of empowering candidates to show their skills and potential. I encourage our customers to use assessments that somewhat resemble on-the-job skills. I don’t like the phrasing “real work” anymore. An assessment shouldn’t be unpaid labor, it should be a way for candidates to demonstrate that they can do the job and handle future work thrown at them, and for hiring managers to feel confident extending what are often very high salaries in tech.
With AI, unfortunately, short take-homes (what I prefer as a candidate, using my own tools and editor) are becoming harder to maintain as a fair signal due to AI assistance. I’ve seen companies move back to onsite, and competitors deploy all kinds of proctoring and invasive monitoring.
The perfect solution, in my view, would be an assessment where the candidate feels relaxed and able to perform at their best, with their own editor and configuration, knowing that every other candidate in the pool has the same constraints in terms of time and tooling. It’s a tough problem to solve. I think about it daily and have not come up with a solution.
I'm using Proxmox primarily to host a mix of VMs and containers that serve various purposes around my home and some utility software / local dev & testing servers. HomeAssistant, Plex, Pi-hole, Monica. Backup once a day to Google Drive.
I've found Raspberry Pis to be too unreliable for my homelab needs due to constant crashing and random performance drops. Instead, I bought two used 2012 Mac Minis with 16GB RAM, 1TB on eBay for about $120 each. I installed Proxmox on them, and it’s been a much better experience. It's nice not having to deal with ARM. They look decent and, most importantly, are super quiet. So can even use them in the living room. Absolutely zero issues with this setup—boring hardware wins.
I think they did a pivot to LLM phone calls? I've tried their library the other day and it works quite well. It even has the "interrupt feature" that is being talked about a few threads up. Supports a ton of backends for transcribe/voice/LLM.
For business class, its often less expensive to buy airline miles than buying a cash ticket. Consider Lufthansa and their Miles & More program. You can currently buy 160k miles for $2,010 (1.26¢ per mile). A return business class flight from the West Coast to Germany is 156k miles and ~$100 in fees. The same flight is usually at least $5,000 for the cash ticket. In my experience, Miles and more tickets are superior to cash tickets, as they can be cancelled and changed for free.
There is availability from SF to Munich on the 25th for 78k miles - go enjoy Oktoberfest!
Absolutely, it's not the most inexpensive option out there. My strategy is, that I only initiate a connection once an agreement with an enterprise client is officially signed. This way, I'm not investing upfront without the assurance of a paying customer. I find this approach works quite well for my situation.
I've been using WorkOS for my SaaS products and I'm pretty satisfied with it. It's straightforward to set up Google/Microsoft/MagicLink (free), including staging and production environments. The best part, though, is that it lets my enterprise customers configure their own SAML/OpenID Connect IDPs.
I get charged per connection, so I just pass that cost onto my customers. As a solo developer, I'd have a hard time supporting so many IDPs with a unified API without this.
They did recently hike up the cost per connection, but they're giving a one-year extension at the old rates, which seems fair.
On another note, isn't it just fantastic that Amazon made a pinky-swear "promise" to not use patient data it acquired with (Alphabet-backed) OneMedical? I mean, what could possibly go wrong with such an ironclad guarantee? It's not like Amazon has a history of exploiting user data for profit or anything. I feel so much better knowing that our medical information is in such trustworthy hands!