I think it would be hard to argue that working all together in an office - like before covid - isn't more productive. In-person meetings are easier. Whiteboarding and brainstorming is way easier. Spontaneous conversations are easier. Helping junior people is easier. People actually pay attention in meetings. You get to know colleagues better.
Have you ever had to "now click on the left... no, up a bit. No go back you were just there. It's the third one from th... I'll just paste the link in chat" when you were standing next to someone's desk? No.
The only benefits of working from home are:
1. No commute.
2. Can do life stuff (we finally have a solution to the dumb problem that shops etc. are only open when people are at work).
3. The company doesn't need to spend money on offices.
The first two are huge bonuses for employees, but the company doesn't give a shit about them. At best they care about paying for offices, but that's pretty minor (especially when they've already paid for them and they're sitting there empty).
I entirely agree on the benefits of in person work. I personally don't enjoy the remoteness of remote work, but the time, money, and sanity cost of commuting still ends up making me prefer working remotely overall.
If I could live in the same building as my job, I would.
> Have you ever had to "now click on the left... no, up a bit. No go back you were just there. It's the third one from th... I'll just paste the link in chat" when you were standing next to someone's desk?
Yes.
I resolve it the same way in person as I do remotely, either by pointing at the screen, asking for (and receiving) control of the mouse and keyboard, or both. If you don't know how to use the features of your screen-sharing software that permit this, you need to learn them. If your screen-sharing software does not have these features, urge your company to switch to one that does (or just switch to one that does, if your company policies permit).
> People actually pay attention in meetings.
I've seen no difference in overall meeting attentiveness in the switch to WFH. Perhaps I've been blessed to work with unusually competent and mindful people.
> Spontaneous conversations are easier. ... You get to know colleagues better.
I don't, no. If anything, spontaneous conversations are way easier remotely, as most of the "water cooler" conversations in the office happen in large "off-topic" channels. This is really good for knowing what's going on at the company, because it massively expands the audience for chatter, gossip, and unofficial news from just a handful of people to the entire company. Plus, it's async communications, which means that you don't miss out just because you weren't there.
> Whiteboarding ... is way easier.
This is true. Most folks don't have cameras hooked up to their whiteboards (or even have whiteboards!).
> [B]rainstorming is way easier.
Disagree. Virtual post-it notes and a shared text document of some kind, plus a group voice call works great.
> I resolve it the same way in person as I do remotely, either by pointing at the screen, asking for (and receiving) control of the mouse and keyboard, or both. If you don't know how to use the features of your screen-sharing software that permit this, you need to learn them.
We used Google Meet, and recently switched to Teams (on Linux). Neither support either feature. I think it's impossible to do it on the web because there's no screen sharing web API (due to security fears) and no overlay drawing API either (again presumably due to security fears).
You need a native app, and Google don't make one, and Teams doesn't support Linux.
> If your screen-sharing software does not have these features, urge your company to switch to one that does (or just switch to one that does, if your company policies permit).
Should you choose to embark on it, I sincerely wish you the best of luck in your quest to get collaboration software that's not crippled by platform limitations.
Having said that...
> I think it's impossible to do it on the web because there's no screen sharing web API...
Weirdly, this isn't true. Discord claims to be able to share your screen even in the web browser version. From the "How To Stream In A Discord Channel" section of [0]:
> Info: Currently audio can only be captured by the Windows desktop, MacOS desktop, Chrome browser, and mobile clients. Unfortunately, audio sharing is unavailable on Linux.
I notice a conspicuous absence of "this doesn't work in a web browser" caveats for any of the other screen-sharing things that Discord does. The only caveats I notice are for audio capture on Firefox and on Linux in general.
And yes, while the Info box at the top of the article says that it covers how to do this in the Desktop Client, I tested screen sharing a moment ago in the latest version of Firefox on Linux. It works just fine. When I choose to share my screen in Discord, the web browser pops up a dialog asking me if I'd like to share my screen, and -if so- would I select which window (or the entire screen) to share?
> ...and no overlay drawing API either...
I don't know how one would handle the "permit others to draw on top of the shared window" mechanism, but there's little doubt in my mind that something can be cooked up that's clever and not too offputting to use... assuming that there's not something already available to be used that neither of us are aware of.
I completely agree, but just as the company couldn't care less about the employee's benefits from WFH, I couldn't care less about the company's benefits from WFO.
Most organizations are so inefficient, that all of this is barely a factor - whether or not work is getting done at 1% of the rate of what it should be getting done, or 1.3% isn't really a dealbreaker for the company.
During the 3 days I wfh, I get the most work done. I can focus and organize my day around executing a plan.
During the 2 days I'm in the office, I can get answers from people much quicker. Some people (new hires in particular) don't know how to describe their problem, or they're just really bad at it. My solution to endless teams convos is to just say "I'll head over to your desk" and then we work it out in person.
I think working with people in person can be very powerful. Is it essential though? No. Most corps don't even bother though. And most managers are bad at management. Working entirely wfh requires good managers with actual project management skills. Most corps are unwilling to train or prioritize hiring for that.
> My solution to endless teams convos is to just say "I'll head over to your desk" and then we work it out in person.
A good replacement for this is a voice or video call with screen sharing. As a bonus, you can rope in folks who are at another site just as easily as you can the first participant. Need to see notes on paper or a whiteboard? Pay the one-time cost to up another camera.
If the folks you work with don't have a "Always respond quickly to urgent messaging requests" habit, then they'll need to develop that. But, IMO, not having that habit is roughly the same as being the type that is rarely at their desk (whether because he's off helping others, or because he prefers to work on a laptop somewhere else in the building).
I find the progress is much faster at the office because me and my coworkers can more easily consult with each other on both major and minor stuff. On chat wfh they might answer immediately or 4 hrs from now. so, I can’t get in a semi flow because of long interruptions waiting for answers. also chat is slow and annoying.
I know someone is going to chime in a claim those conversations break flow. That might be true for some particular deep problems but I’ve never seen it affect any team I’ve been on and I have 40 years of work experience. Not saying my experience fits everyone else but I’ve seen no real evidence of it being an actual problem.
I don’t know the reason, nor do I have proof of anything. But: to me this is a great time to consider Occam’s Razor.
Executives seem to (mostly) universally want people to RTO. Why would they?
They obviously have lots of data. If it was bad for productivity, why would they do it?
Answers seem to be things like “power trip” or “need to justify real estate”. I’m pretty sure most companies would save money by giving up their leases. Maybe they are all having power trips, but irrational behavior from leaders won’t win out in the long run.
My observation from my time is that, likely: some people are really good at getting stuff done at home. But most probably get less done. And I suspect the leaders find this in their data.
I'm sympathetic to your reasoning, and I actually think your conclusion about how people wfh is probably true (a minority finding it wildly more productive) but the problem I see is that most execs don't understand what a lot of their employees do on a day to day basis, nor do I think they could properly understand the data even if they had it as a result. And, if they did have it, why don't they show it if the conclusions are so self evident?
In reality I think it's much simpler: the work that executives (and some upper level managers) do relies on a certain amount of theater. They need to be seen (in person) and they to be seen doing work (in person). That's part of the "deal" with being an exec --- you need to be able to act the part. Then, they just assume that's how it should or needs to be for everyone else.
The interesting thing is that "Darwinism" will sort it out for us in the long run. If the execs are right: more work gets done in the office, those companies will do better; if people are happier and more productive at home, those companies will tend to do better.
There is a part of the population that are opportunistic freeloaders. People who don't have intrinsic motivation to do their work but do it to avoid the consequences.
I think it's a pretty big part of the population! And I also think that, while managers are likely to have encountered them, reaching a software development role (and then choosing to spend your spare time commenting on HN) filters out a lot of them. So there is a mismatch where managers want to defend against the freeloaders while the employees commenting here can't accept the lack of trust inherent in the mandate and won't countenance the micro-management needed to prove that trust would be earned.
I think it also partly explains why we've failed to see remote companies gain a competitive advantage in hiring. They have adverse selection for active freeloaders.
I'm really not sure where 1 to 3 come from? 1 and 2 are about spending time with people. 3 can't also be true, just look at all the innovations that OSS has created (at least 90 perc. remote collaboration).
Number 4 is a real problem and a symptom of the biggest problem imo, which is that it's hard to build trust online/remotely.
Things that are inherently "about spending time with people" are very likely to be more effective in a higher bandwidth, more natural setting (read: in-person).
On the spectrum from postal mail, e-mail, text/instant messaging, phone calls, video calls, in-person discussions, the quality of "spending time with people" is increasing as you go farther down the list.
It is true, and spending time with people over video calls or text chats is a very low-bandwidth, rather thankless job which is generally demotivating long-term. Not that many people are actually willing to channel their emotions for hours daily into pixels on screen with no actual emotional human connection. This means that juniors remain isolated, senior employees feel that tutoring juniors is a waste of their time (which is true in this setup), and good managers either stop caring or burn out.
> 3 can't also be true, just look at all the innovations that OSS has created
Innovating in things you deeply care about is one thing, innovating for yet another work project that brings shareholder value is another. The only recipe that has worked for the latter so far has been "put multiple smart people in one room".
I personally feel home office makes jobs feel much more like jobs. You have no social relation to your coworkers and it's very hard to build one remotely. Most people barely communicate in zoom meetings beyond the minimum what is required of them.
Good office jobs can feel fun, building something with like minded people and building a social relationship at the same time with them.
Many people are not looking for that, many people just want a job that makes money and let's them log off at a set time. I personally don't enjoy that type of job and feel it's not sustainable over decades. I need some fun in my job.
Sure, I'm on both sides depending on the situation so I can help here.
"Collaboration" and "communication" are definitely issues, as employers state openly, but not for the reasons you think.
I worked in a fully remote company for the last three years, I selected people who I knew would be independently motivated and who knew how to communicate effectively in an async manner.
Yet, there were issues still in communication that would have been solved with being able to hear someones tone and body language. Little issues spiralled out of control- and lack of trust crept in when people were quietly doing work that had very small visibility (but was important).
I'm not saying remote work was the reason, but remote work exacerbated certain particular issues in this dimension- eventually the founders (who had lost trust) started laying off people they perceived as not important, despite them being fundamental to any future in which the project would be successful.
You'll also notice that the companies that care the most about RTO do not invest in communication in other ways, they'll probably use awful communication software like Teams, WebEx - or in marginally better cases: Zoom.. and since people don't like to interface with these tools (and they're poor even when you do): there can be a feeling of being out of the loop from upper management as they don't know what's going on.. they just see radio silence.
The failed Windsurf bid and this makes me think OpenAI feels they need to focus more on the coding agent use case.
Still thinking about the endgame. Its not obvious to me if OpenAI/Anthropic will become competitors to coding startups like Cursor or continue to be model providers.
I think the endgame is a shift toward a platform of services that tightly bind users to a single LLM provider.
Right now, many small startups are essentially just thin wrappers around ChatGPT. Once it becomes clear which ideas and solutions gain real traction, providers like OpenAI/Anthropic can simply roll out those features natively removing any need for a third party.
In a sense, a lot of what happened with the mobile market. For example, there's no need for a QR scanner or document scanner app anymore, if your phone starts to offer it natively.
Does anyone else remember how we used to have flash light apps all over the playstore and how they quickly varnished once the feature was implement natively?
Almost everything in ios since ios1 has been a sort of in house copy of a jailbroken tweak. I feel like that is just how these huge companies actually innovate featuresets now. No one has agency to drive change from within, so they see what the community has built and is using and shamelessly copy it as it. Easier to sell to management that you should steal an already bright idea than to try and prove your own novel idea is worth pursuing.
This seems correct to me. I'm unusually appreciative of vertical integration (life-long Apple user, etc.), and I can already feel the vendor lock-in tightening. I have no need for anything other than my ChatGPT subscription, and adding other tools appears to offer marginal gain at double the cost.
This type of bundling appears to be one of the strongest forces in the economy today, and I think comes about consistently due to a confluence of efficiencies of scale, coordination, and second-order effects of prestige (being able to hire and pay large numbers of outlier high performing employees, etc.)
I've learned not to bet against it, except in niche areas.
Neither type has agents in their DNA. The IDE agent companies are dragging an enormous vestigial appendage, and the the scientists at frontier labs are stuck in the next token prediction mindset. This needs a systems engineering approach with online learning from feedback and high throughput optimization experiments run end to end.
> Its not obvious to me if OpenAI/Anthropic will become competitors to coding startups like Cursor or continue to be model providers.
They’ll do both: continue to be model providers while also leveraging their position as model providers to own as many of the valauble markets in which models are used as possible. Kind of like Amazon and its role as both infrastructure provider and direct competitors to other sellers (on the shopping/logistics side) and SaaS vendors (on the AWS side).
I really don't think this is a good idea. All the negative comments seem to have been from people who almost treated 4o as a friend rather than a tool. I don't think encouraging that direction is good in any way.
My job uses tiny models to determine escalations to bigger models. The tiny model provides a label and if it’s high confidence, we escalate to ChatGPT confirm.
I also want to try this with language detection. Existing open source ML models have weaknesses for mixed language, length of text, or domain limitations in the underlying text (like trained on bible translations).
I’m not sure what I’d use them for, except maybe tag generation? Encoders of this size usually outperform by a wide margin on tasks they would overlap with.
I'm making an app where literally all I want to do with an LLM is generate tags. This model has failed with flying colours, literally takes forever to parse anything and doesn't follow instructions.
Edit - I should add, currently the model I'm using is Gemini Flash Lite through the Gemini API. It's a really good combo of fast, follows instructions, gives correct results for what I want and cost-effective. I still would love a small open model that can run on edge though.
Use a large model to generate outputs that you're happy with, then use the inputs (including the same prompt) and outputs to teach 270M what you want from it.
We're currently running ~30 Llama 3.1 models each with a different fine-tuned LoRa layer for their specific tasks. There was some initial pain as we refined the prompts but have been stable and happy for a while.
Since the Qwen3 0.6B model came out we've been training those. We can't quite compare apples-to-apples, we have a better deeper training data-set from pathological cases and exceptional cases that came out of our production environment. Those right now are looking like they're about at parity with our existing stack for quality and quite a bit faster.
I'm going to try and run through one of our training regimen with this model and see how it compares. Not quite running models this small yet, but it wouldn't surprise me if we could.
I built a reranker for a RAG system using a tiny model. After the candidate generation (i.e. vector search + BM25) and business logic filters/ACL checks the remainder of the chunks went through a model that given the user query told you whether or not the chunk was really relevant. That hit production, but once the context size of models grew that particular piece was discarded as passing everything yielded better results and prices (the fact that prices of input tokens went down also played a role I am sure).
I agree that FastAPI is not the best but I would not jump to another framework just because of an article.
For FastAPI I know that I can go to [1] and see what a good FastAPI code base looks like. These days even Airflow is on FastAPI but haven't looked at the codebase.
For me to jump on Litestar, I would like to see a reference codebase to learn best practices. Otherwise its one more framework whose quirks I have to get comfortable with.
Use pymupdf to extract the PDF text. Hell, run that nasty business through an LLM as step-2 to get a beautiful clean markdown version of the text. Lord knows the PDF format is horribly complex!
We OCR them with an LLM into markdown. Super expensive and slow but way more reliable than trying to decode insanely structured PDFs that users upload, which often include pages that are images of the text, or diagrams and figures that need to be read.
Really depends on your scale and speed requirements.
Just to clarify, this requires both the client and the debugging script to be running 3.14? So I cannot have a script running an older version of Python but use a Python 3.14 debug script to attach to the running script?
So Python 3.14 introduces this functionality. The process you want to debug needs to be on at least Python 3.14.
If you want to "just" use sys.remote_exec, you will need to use that from a matching Python version. If you dive into the implementation (check out the PEP), you could most definitely write a C or Rust program that implements remote_exec and "does the thing".
The protocol is basically "tell the program to read a python script from this file". That python script will be executed in the running program's environment, but you don't need for the software on the other side to be a python program!
sys.remote_exec itself only targets same Python versions though.
Zed is not a great modal editor, modality is a strapped on afterthought. At least that the impression the vim keybinding give me. IMO the best chance I see for a fast modal editor that works out of the box is Helix.
When did you last try Zed? If it was recently then that hasn’t quite matched my experience, for me vim keybindings work quite well. But I did think the same when I first used Zed over a year ago.
I tried it a couple days ago. Lost me right off with the poor `:e` experience; popup box with no completion, fuzzy-matching, sub-directory support. Vim is a file oriented editor, not a project oriented one, so good file operations are table stakes.
I don’t quite get what Zed’s emulation of :e has to do with being a modal editor. It’s a completely different editor so of course the interface to interact with files is going to be different, but in my experience Zed’s implementation of modality when it comes to its Vim mode is extremely good. Even better than the already excellent NeoVintageous for Sublime Text.
Fair point. I guess the problem is with their (Neo)Vim keybindings, not so much with the modal-editing part. I included it as vim is the most popular modal editor and its keybindings are what are lacking.
Are there any modal editors that are not file-based? Maybe it would be better to say this is a problem with its file-based editor emulation?
I at least really like that ctrl-w allows switching between panels as well as editors, something VS code doesn’t do (or at least, you can switch to the explorer but not back?) and is a major papercut IMO
Huh? Zed's vim emulator is one of the best I've used in an IDE. Saying it's an afterthought feels disingenuous. I'm pretty sure some of the core zed devs are big vim enthusiasts.
I don’t want to install 5-10 plugins with a hundreds of config lines just to make it usable. I prefer zed with 0 extensions and vim motions which are built in.
Not really, it installs even more plugins. Some of them are unnecessary.
Ideally, every time you update or install a plugin, it must be reviewed.
I just don't wanna deal with it and trust plugin developers.
I already got bitten by an infected Python library (that used to be legit). I treat plugins the same since they can be used in supply chain attacks.
You have to trust VS Code plugins too, along with trusting MS not using your code to improve their AI suggestions, or doing a sudden rug pull. All three happened if I remember correctly.
You can easily enable/disable plugins in LazyVim, and they are simple Lua Script/VIM scripts easy to review. I don't think VSCode would be that easy.
And even with more plugins, nvim/vim will be much more performant then VSCode.
I switched to Helix for the same reason. Though I'm thinking of re-trying with one of the lighter weight pre-configured versions. Astronvim seems like the best one after a bit of weekend research.
What I would really like to see is arguments from the other side. Can someone steelman RTO. Preferably with evidence, anecdotal or otherwise.
reply