jillesvangurp

Open models still need to run somewhere. If that's not your machine, expect to pay or watch some ads.

Good that they are integrating that into the main browser.

For me the big use case is having different Google accounts active in one window. Work and private. And when I was doing freelance projects, I sometimes had 3 or 4 different Google accounts active. Plus assorted stuff tied to those accounts. Using different colors for each makes it easy to tell apart the tabs.

Another use case is if you are testing the same app for different customers with different credentials. I do this a lot and I have dedicated containers for each customer so I can easily switch with out having to log out and in a lot.

And of course some companies that can't be trusted with cookies, like Facebook, get isolated. I open that container less and less these days. But I still check every few months. I also keep my bank's website in its own container just on principle.

Bots in team chat is not a bad idea. I've been experimenting with that for a few months.

We started with Slack. This kind of works but dialing in the gazillion of permissions needed to end up with something that works is a bit painful. And you have to do this over and over again for each new bot.

So, we started experimenting with self hosted alternatives. First we tried Matrix. It's nice but a bit too strict on end to end encryption which gets in the way if you actually want to share information with bots. So, a few weeks ago we switched to Zulip. That's pretty straightforward to setup as well. It's very easy to create bot users with both and create some automations. We did some with Openclaw and I ended up replacing those with something coded around the haystack framework, which is a bit less of a messy basket case than Openclaw.

And as I discovered after I installed it, the Zulip leadership actually just got hired by Anthropic. It seems Jack Dorsey beat them to market making some announcements but I imagine that Anthropic might have very similar plans.

IMHO team level use cases for agents actually make a lot of sense. Companies are run by groups of people and the larger those groups the more team overhead you get. Perfect for optimizing with AI. And if you have people doing a lot of stuff with AI, that only adds to the need for them to communicate what they are doing and coordinating with others. Doing that out in the open in some shared channel makes a lot of sense. And having some shared team guard rails also makes sense. And if you think about more complex processes with handoffs between people (and possibly some agents), a team chat is a good place to do those as well.

I don't think it's that black and white. OpenAI and Anthropic are building valuable tools on top of their models. You get a very rough and much less polished version of that with open source tools and and open source models. And you still need inference infrastructure to run those. But at this point most of the competition is in the tool ecosystem, not the models. And while there are plenty of people toying with things like opencode there's a clear pecking order emerging where Codex and Claude Code/Cowork are generally considered the top choices before tools like MS Copilot, Gemini, and then a rapidly shrinking long tail of alternatives to those.

In the end what companies pay for is not tokens but results. A DIY kit of models, mac minis or whatever, and a bunch of poorly integrated OSS tools doesn't solve their problem. For the same reason, people use Office 365 rather than running Libreoffice. And for the same reason things like AWS dominate the market rather than people DIYing their infrastructure together themselves. Most of the money is in polished turn key solutions. Which is what Anthropic and OpenAI offer.

The juicy market here is the enterprise market. That's mostly business users, not programmers. They'll be hooking up all their SAAS tools (which they also over pay for), and other stuff. They'll be paying for boring things like data residency, compliance, etc. And they need access to reliable infrastructure to run all this stuff. They'll want this shit to just work and not to be dealing with a lot of poorly integrated stuff.

Most of the billions invested are being sunk into infrastructure, chip design, and access to resources (land, water, energy) needed to run data centers. A handful of companies now own most of that infrastructure and they also happen to have the top models, researchers, the best tools, and warm customer relations. And they sell access via very convenient subscriptions with high enough limits that people don't have to worry about things like token cost. The game here is recurring revenue from customers that like predictable pricing, reliable quality of service, and iron clad compliance and data security & residency, and quality guarantees. These companies don't want to be chasing model quality and have to upgrade their entire company every few weeks. They want continuity and predictability. Mostly they just pay Anthropic, OpenAI, MS, or Google to take care of this for them. There might be some niche EU players that become a bit bigger. But I don't see a large scale switching to Chinese suppliers for a full polished alternative. The Chinese might give away their models. But I don't think they'll be generating a lot of revenue.

And if you want to run your own models, you'll still need infrastructure to run it. These four companies together with the usual cloud giants control most of that and as well of the supply of resources (chips, data centers, energy, etc.) in the EU and US markets. There's going to be a long tail of self hosted and gobbled together stuff but it's going to be a much rougher experience for end users and it won't likely be most of the market any time soon.

Works fine for me. I also took a zoom call from Firefox just yesterday (153.0b13 on mac). So, I think this is false alarm.

I also see no good reason why Zoom would want to restrict this. Likewise, I don't see them dropping support for Safari either (works fine as well).

Maybe they limit things on mobile? But on IOS they'd be stuck having to support the Safari browser engine anyway as neither the Chrome nor Firefox browser engines are used on that platform (both browser delegate to Apple's browser engine there).

I just made chat gpt list all the key indications that this article was LLM generated. I'll spare you the details, it's a very long list.

It's actually possible to counter some of this stuff with skills and guard rails. The author clearly has not mastered how to do that and sound more authentic.

It's all a bit stunted, cringe, long winded, repetitive, etc.

It seems Google is repeating it's mistakes of chat apps (they had gazillions of those at some point) with AI. There are just way too many tools that they are pushing. And none of them address the glaring issue that most of their users end up using Codex or Claude Cowork instead. Google is failing hard at convincing their own users that they are as good. They seem to have a lot of fragmented efforts that don't really add up to good enough.

I grew up in the eighties. Tapes, records, cds, etc. I've seen it all. I haven't used a device that can play any of these in over two decades. I have more music at my fingertips in Spotify than I ever had on these. Love it.

I was a teenager in a small provincial town in the Netherlands. It had two record stores. They were small. There was also the library with a small collection of decent cds that you could "borrow" (copy to tape). If it wasn't on the radio or in those stores, or in the library, I had no access to it. That was how I built my music collection. Buying albums properly was expensive. I only had about 10 or so. I had a few dozen tapes. Maybe 60 Albums or so and some mix tapes. That was it. Compared to what I have on Spotify, that's nothing.

There's all this BS. how everything was better on "hifi" equipment of the 1980s. The reality of course was that most poor teenagers did not have access to either decent equipment or decent music. I tracked down most of what I had in physical form on Spotify ages ago. And I even listen to some of it regularly. Sounds great with decent modern headphones. I'm perfectly happy with my Sennheiser bluetooth headphones. It probably sounds infinitely better than what I had back in the day. I alsp have a much wider selection of music. Anything I want really.

And of course I'm now in my fifties and my hearing isn't what it used to be in terms of high frequencies. I'm about average in hearing loss (i.e. not going deaf) but any absurdly priced headphones would probably be wasted on me. That's the irony of aging audiophiles, right when they can afford all the right gear, their hearing starts deteriorating. You can't compensate for that with gold plated plugs or any of the other nonsense people are into. Records and analog amplifiers sound nice mainly because of the sound distortion and compression. This isn't all that hard to emulate with digital processing. The pop music of the nineteen eighties was optimized for the cheap FM radios that most people had. That's why a lot of albums got remastered later. The original sound is not actually all that great on modern equipment.

The way around this is the people that listened to this also listened to that feature. That's the main way I discover new music these days. I find something new that I like and then I explore whatever it is that people listen to that also listened to that song. If you don't do that, the recommendations are basically more of the same shit. All the B-singles of an artist you like some songs of. Or worse: "you're old, here's some old music for you!". Or even worse, "your ip address is in Germany! You know you are a closet German! Here's some German hoompahpah music for you!".

Amazon and Youtube are equally useless when it comes to recommendations. All the machine learning talent in the world and they are utterly useless. I clicked a young ones clip on youtube a few weeks ago. Now my recommendations are 40% more f*ing young ones clips.

But randomly clicking stuff on Spotify reminds me of 25 years ago where you'd randomly download some shit and then listen to it. I also miss the art of a well produced album. I can't listen to individual songs of a good album. I have to play it beginning to end. I hate all the bonus tracks that Spotify slaps on albums. The whole point of the last track on Dark side of the Moon is the fade out to silence at the end. But that's just me.

I wish they would stop breaking my playlists by randomly breaking links to songs when they get replaced or re-imported. Seems I have to hunt down replacement tracks for 10-15% of my carefully curated playlists every year or so. Usually they are still there. But in some cases entire albums disappear. All the endless remasters, best off collections, etc. that they keep churning out result in endless breakage. How hard can it be to automatically replace those songs with the exact same recording on a different album?!

There are a few things going on at the same time that mean that electricity cost might actually go down.

The reasons for that are complex but have to do with how electricity pricing works. In many markets the price includes a lot of taxes, fixed cost for providers and infrastructure. Generation is only a minor cost. And on top of that the prices are set in a way that isn't really that flexible.

Infrastructure utilization is a very important here. Grid operators are very conservative with their infrastructure. They want to ensure there's enough to handle the worst case. That means there are a lot of assets that are nowhere near 100% utilized (e.g. cables and long distance transmission). It also means they are very inflexible serving new demand like data centers.

Adding batteries as energy buffers enables a lot better utilization of all these assets. That enables more revenue for the same infrastructure cost. Electricity prices can actually go down if you do that right. With renewables, there is very low marginal cost for generation. It's all infrastructure cost. Anything that improves infrastructure utilization enables more customers to have power that then share the infrastructure cost.

Data centers that are currently powered by things like on site gas turbines are not being very cost efficient. There's an obvious incentive for hyper scalers to invest in infrastructure that will lower their cost. They have access to many billions. They are spending on anything that will get them energy. They are desperate to spend. And they are completely bottle necked on grid operators that are being very conservative. Hence the expensive side hustle with gas turbines. There's a big opportunity here for massive investment in better grid infrastructure. That wouldn't necessarily be payed for by consumers. But they would still benefit from better infrastructure.

The key is unlocking these investments to happen.

SaaS partially took the place of bespoke projects that people were doing before. They never stopped doing those of course and there were also off the shelf packages that people bought before SaaS.

AI lowers the cost of creating bespoke software that competes with both. Instead of buying a one size fits all thing that half does what you need, you can now have a thing that is a bit better suited to your needs. There will be a lot more demand for those things. A lot of these things are going to require deep domain knowledge and some system thinking skills.

This is still hard enough to do well that a lot of the creation work will be outsourced to professionals. Even if that involves the use of AI prompting. Maybe after naive attempts to do it in house fail. My hunch is that there will be a lot of growth for those that can do these types of projects efficiently and that it might more than offset the job losses in the SaaS sector.

There are a lot of of companies that are still under using software. There never was any good SaaS that fit their needs and they lacked the skills to do it themselves. When you lower the cost of something (creating software) the market usually grows. A lot of things that were previously not feasible are now doable.

You'd pretty much have to self host on premise or in a fully certified environment for classified stuff like that. That will limit your options of course. But some of the OSS models might still be OK for this.

The article seems to be very unspecific about what it is this company does that is so different. It also steps over the fact that there are already quite a few companies active in the US, EU, and China that are recycling batteries. Nor is the cited percentage that remarkable. That's ballpark what competitors are achieving as well. Probably a bit more. 10% lithium is a lot of lithium to not recover. Most natural deposits of lithium have very low concentrations of it.

The main thing actually holding back the recycling industry is the lack of batteries that need recycling, not the lack of technology needed to recycle them. Most of the batteries produced in the last ten years are still being used. And quite a few might head for a second life in storage for another decade or so. It's probably going to be another decade before recycling hits a scale where it becomes a significant and lucrative source of valuable raw materials.

And as others mentioned, it's not just about recycling the lithium in batteries. It's not like cobalt, nickel, copper, graphite, etc. end up on the trash heap.

We gave it a few months try. My conclusion was that it did interesting things but is designed completely wrong from the ground up which means it's got a lot of moving parts that are part of the code base that can break for all sorts of reasons.

As a learning exercise it was pretty nice. But since then the plugin/connector ecosystem for things like Claude Code/Cowork and ChatGpt/Codex has first come into existence and then matured to the point where they do most of what made OpenClaw interesting before that was the case. But in a way that can be shipped to lots of users. That wasn't true when I started using it. We played with it for a bit but in the end the code base is a mess, random shit breaks every time you update it, and you end up doing very dodgy shit with it that no responsible CIO would want to sign off on.

The current batch of tools from Anthropic and OpenAI is pretty solid but there are still lots of feature gaps, UX issues, security challenges, etc.

I think a lot of these tools will get into the enterprise in exactly the same way that usb sticks, MS Office, smart phones and other consumer tech got into the enterprise: employees will bring them and use them. Some bosses will tell them off and then make an exception for themselves. Because the promise of not having to do monkey work that can now be automated is unbelievably tempting if you need to do lots of that work. Even if this stuff only half delivers on some of the promises. So, my guess is that this could go fairly quickly. I already see a lot of non technical people that are pretty clued in to things like Claude Cowork. There will be some rogue early adoption followed by more enterprise appropriate solutions. That's already happening.

Here in Germany, there are a few German AI companies that work with big conservative German companies and the public sector (e.g. Langdock and Deepset). The type of organization where privacy and data security concerns weigh heavily. These companies can work with OSS models but OpenAI can do proper data residency in Germany as well if you talk to them. Azure, AWS, and others have very acceptable options. And there's a whole ecosystem of companies that are building on top of that.

You seem to assume that people still have an inherent advantage here. Mostly those sectors compensate for well documented human failures with very rigid and expensive processes and testing. And of course despite that, stuff still goes wrong occasionally.

I actually think AI based automation is going to be a key enabler in those kinds of strict environments as well. Recent work on identifying e.g. security bugs seems to be resulting in a lot of improvement that has somehow escaped decades of human scrutiny. With good quality harnesses (manual or automated), I don't see any big objection against using generated code here.

Sensible policy. I think with mainstream news publications now obviously using LLMs in their day to day workflows, it's going to be hard to take a purist stance here. Some do this more responsibly than others. But I don't think it's necessarily a bad thing if it is done responsibly. It's only a problem if it is done poorly.

Poor writing is not a new thing, of course. Most of the moderation mechanisms that it uses were perfected a quarter century ago when sites like Slashdot were popular as a defense mechanism against bad user behavior. Bad user behavior impacts commenting, article submissions, and moderating itself. While bad users now have AI to abuse, the problem of a large volume of low quality content is is largely the same. The moderation mechanisms end up targeting the problem at the source: identifying good and bad users. So, the same amount of bad users generating a lot more garbage isn't that big of a deal. Getting good karma still is a lot of work and it makes identifying all the garbage created by users without that relatively straightforward.

Using LLMs to tag, flag and filter content might not be a bad thing to experiment with. There are a lot of low quality AI generated opinion pieces that somehow make it to the front page. Same for political and controversial stuff, which of course is against the HN guidelines for content. Auto flagging things that obviously violate guidelines should not be that hard. It's just a matter of having good guard rails. @dang might actually already be doing that. I know I would be if staying on top of piles of generated garbage was part of my job description. It might also be done to give good content a little boost.

The new articles section has a very low signal to noise ratio currently and the window for good content to make it past that is very short. Often articles on the front page will have many duplicate submissions that never made it past that. IMHO duplicate submissions should just count as upvotes on the original. Auto de-duplicating based on canonical URL should not be that hard.

The relevant question is who is going to pay you to write code manually. It's something that's increasingly hard to justify. The answer is of course that people are not paying for code to be written or generated but for some problem to be solved. Whoever does that with the least amount of drama and cost gets the business. And using AI tools just allows for compressing the timelines a bit in a way that is hard to ignore.

Most code out there isn't all that great. I've been in this industry since the nineties. There are a lot of not so great software engineers doing mediocre work. People are romanticizing how great and magnificent their code is (or used to be). The reality of manually crafted code is of course a lot less flattering. Many code bases become hard to maintain over time and are riddled with bugs. And it's not like sloppy code is a new thing. Poorly executed software projects have been very common for a long time. That's not going to go away.

People seem to have a poor understanding of just how much space there is up there. It's just very empty up there. And these things are in precisely controlled orbits that are well documented, etc. Even if you simplify your thinking of orbits to a 2D (square area), it's a lot of space.

But that would be a mistake of course. Low earth orbit is three dimensional. Star Link uses several altitude bands of about 20-30km each. It's 330-360km for the v3 satellites. The volume of that is about 17 billion cubic kilometers. About 13x the volume of all the water in the oceans. Accidental collisions are not going to be a frequent thing. These things are going to be many kilometers apart.

Many people have never seen that properly due to light pollution.

The star link network is actually remarkably cost effective in getting internet access to rural areas. There's a reason that these areas still have poor connectivity: it's just not cost effective for anyone to build land based infrastructure there.

SpaceX spend a few billions on StarLink. But if you look at how much network operators have spent over the years on cables, base stations, etc. it's not all that much for a network that offers high bandwidth access all over the planet.

Adding 100K more satellites is going to make Star Link a direct competitor to many of these operators.

I've barely touched Intellij for the last half year or so. I rarely edit code manually at this point. I also did not renew my subscription and am back on the community edition.

I've noticed that my preferences for tools and languages is shifting as well. I'm happy to work with stuff that I previously would have not touched now because it would take me too long to get up to speed with languages, frameworks, etc. That stuff no longer blocks me from being productive. I still care about code quality, good design, etc. but a lot of that stuff doesn't require me to micro manage a code base. In the rare case I want to open something in an editor, I use vs code. I've removed a lot of the plugins in that as I'm not really using them any more.

I actually do reach for vi on the command line sometimes. But I've never been very good with it. I know how to do simple edits and save the file. I just never really got into it. I memorized a handful of key bindings somewhere in the nineties and that's it. I know some people that live in this editor and swear by it but for me it's just something that's there by default that is vaguely useful in a pinch if there's nothing else.

100% of the tests passing, on track to be faster and more scalable. That's not a trivial achievement.

If you know your military history, it's a lesson military planners learn with essentially every war that is then forgotten when the next generation of military planners and politicians come along.

Both Russia and the US learned expensive lessons in Afghanistan fairly recently. And yet here they are engaged in conflicts in Ukraine and Iran that don't seem to go as they planned.

People feel threatened by LLMs doing things well that they feel should require their skills and talent.

That's understandable but it's still a bit of a negative emotion that probably isn't very productive. Or very rational. This thread is full of people trying to argue that this can't be any good, shouldn't be any good, and is clearly going to end in tears. And obviously this thing passing tens of thousands of carefully curated tests that accumulated over decades suggests otherwise. It's hard to argue against that.

This probably is going to have some new issues. But it's an impressive achievement.

That's very true. People put up with the many limitations of off the shelf software because it's cheaper, not because it's better. Developing bespoke software solutions is now a lot cheaper than it used to be. So, there are a lot of cases where that now becomes the better option.

Doing in days what used to take months, is a bit of a game changer. Like with past cost reductions, people will underestimate the work and get it wrong. It helps if you know what you are doing rather than just vibe coding things.

But for rewrites, the sunk cost fallacy becomes a lot cheaper. So, that changes how you deal with stuff that clearly isn't living up to expectations. Unceremoniously replacing what wasn't that expensive to begin with might be the cheaper option relative to fixing it.

I've been living in Berlin for about seventeen years. My German isn't great but usable in an emergency. I get by with it. Most work related stuff is English.

Bureaucracy is an annoyance in this country. But the flip side is that if you persist, you'll manage. It's also not something that's necessarily a lot better in other big countries. But Germany could do a lot better by just moving a lot of the key processes online, cutting down on asking for the same information over and over again via paper forms, and speeding up decision processes. That's slowly happening.

With AI translations, doing things in German (or any language) is a lot easier these days for foreigners. Also making sense of the complex processes with AI is helpful. Insistence that everybody should learn German is understandable from a nationalistic point of view. But you get quite far without that. Easier than ever now. Germany could be a bit more accommodating for this.

And the reality in factories, on construction sites, etc. is that you hear a lot of other languages being spoken. Lots of eastern Europeans active in the construction industry, for example. And lots of nurses and doctors from abroad are active in their hospitals. Packages are being delivered by people from India and Pakistan. And of course German companies that sell to foreign companies have to deal with the notion that their customers mostly won't be speaking German. Germany is already a lot more international than it might like to admit.

But it's undeniably true that you need to speak German in order to interact with especially older Germans and their companies. They simply don't speak anything else. Kind of weird because many of them are super dependent on import/export markets and yet they are mortally afraid of having to be in a meeting with non German speakers. I've experienced this several times. However, the baby boom generation is retiring and younger generations are already much more internationally focused. Most younger college educated people here speak English at this point. It's not that much of a problem as it used to be.

And even talking to people is getting easier now that we have AI translations and transcriptions. I've worked my way through a few meetings in Denglish. Ugly, but it works and if you have a shared business goal, people get more flexible.

Germany has been in and out of a recession for several years now and it's working population is on track to shrink and things like its pension and healthcare system are becoming a problem financially. It will need to work smarter to get out of that and that probably is going to require working with people that won't be speaking German from outside of Germany. Easy fix for that recession is just embracing the future. Many of Germany's problems are of its own making and very fixable.

That is indeed a simple case. Imagine using that on your Google drive where you might have lots of variations of the same documents with names such as my-doc-draft.doc, my-doc-outdated.doc and my-doc-final-v0.1.3.doc. Which is the best version? Unless you add logic to control this, even the best vector search will not be able to tell apart outdated/incomplete/inaccurate data from the best data. That's why enterprise document search never really got good. Google's own drive is a good example of a search that isn't great.

A good example from an ecommerce vendor that now trains its own models is photo search. They used an off the shelf model to implement that when testing with a photo of some clothes, instead of getting similar products they got random products that featured the same person modeling completely different clothes. The model they used was biased towards faces rather than clothes.

A lot of these US vendors have data centers in the EU operating under EU law via legal entities in the EU. So, it's not all that black and white. And there are a growing amount of EU based alternatives.

And just to make a counter point, there are also US dependencies on the EU for some things as well. Mobile infrastructure is a good example; mostly comes from Nokia and Ericsson. What was left of US based network equipment makers was merged and acquired in the early 2000s. For example Bell Labs is currently owned by Nokia. It includes bits and bobs that once belonged to companies like Lucent and Motorola.

Another dependency is shipping; the US has very few ship yards left and is looking increasingly to the EU for things like icebreakers and some navy ships. Likewise, ASML the industry leader when it comes to making lithography machines used in chip making is based in the EU as well. And of course a lot of manufacturing uses machines made in e.g. Germany.

IMHO this mutual interdependence is actually a good thing. It stimulates maintaining peaceful relations and engaging in trade. We could use a little more of that. Isolationism didn't lead to anything good last century either.

RAG is a fancy acronym that basically boils down to: let's give ai agents the super power of information retrieval (aka. search) and "augment" the generation with a list of results by adding that to the context.

The narrow interpretation of this is usually some kind of vector search. Which some people naively treat as magic pixie dust that will make search quality amazing without any tuning whatsoever.

This does not actually work all that well beyond really simple use cases. A well tuned traditional search engine can be surprisingly competitive. And I know people that do pretty complicated things with vector search that usually involve training their own models and spending a lot of effort on testing and validating those are any good.

I've been doing stuff with search for a bit over two decades. AI use cases makes information retrieval more relevant than ever. It's a key ingredient to answering questions for complex, proprietary data. And especially when that data is very complex and unstructured, naive approaches tend to have their limitations. In other words, it can pay off to to sit down and do it properly and think about things like data ingestion pipelines, transforming & enriching data, testing search quality, etc. Most of the success of a good search system usually boils down to getting your data right for indexing and optimizing it for how you are going to query your data.

The good news is that with large context windows, precision (best results are at the top) matters a bit less than recall (the search returns what you need when you search for it) these days. You can compensate for imprecise search by just fetching more results. As long as what you needed appears somewhere in the top 500 or so, you'll be fine. The flip side is of course that you end up adding a lot of noise to your context which might throw the LLM off and in general wastes a lot of tokens. That's why precision is still important.

What the article is proposing is post processing imprecise results to filter out the noise with an LLM to compensate for what is basically not a great search implementation. That can work of course (provided your recall doesn't suck). But it's going to add some cost and latency to searches. And usually, agents do multiple searches.

But if your search is so poor, why bother with vector search at all? Especially dense vector search at scale is not cheap. If you are going to fetch lots of results, just use some cheap lexical searches. Sparse vector search might be a good compromise (higher cost to index but similar performance to lexical search).

It's very similar to people that feel compelled to constantly name drop people and companies they've heard off that supposedly did some cool stuff that impressed them. A lot of the tech scene is people just blabbing at each other about who and what they've heard about.