mdasen
I think the first implying the second might be a bit of a stretch. At some point, companies are going to start looking at their AI spending and start negotiating deals. OpenAI is right behind them and Anthropic is around 8x more expensive (Sol 6.1 vs Opus 5.5 based on Artificial Intelligence's cost per task; cost per token is a somewhat useless metric when different models use vastly different numbers of tokens for the same thing). Chinese models like Deepseek 4.1 Flash are cheap, can be self-hosted, and while certainly behind Claude Opus 5.5, are still good.
Lots of companies buy "inferior" products like Microsoft Teams. Over the next 5 years, it seems likely that many models will get really good to the point that people often won't be able to tell any difference. Even if something isn't "definitively" better, companies might say "our engineers can't tell the difference and we're cutting our costs by 85%."
It's not absurd to think that a company would use its own harness or an open source harness and/or put a proxy in to route requests variably. For two months, they have engineers using different models unbeknownst to them and they measure things like how often the engineers have to correct or redirect the model. At the end of two months, they might find little difference in performance and switch to a much lower cost model.
It's not like it'd be crazy hard to analyze workers frustration as they did their jobs and figure out if one model was actually better. They could log all the conversations and it's not like AI can't easily determine things like "no, that's completely wrong" as a phrase showing that the agent was really off base.
I think a big part of the next 5-10 years is going to be about cost/efficiency.
JPEG XL also has some big benefits for many use cases like lossless re-encoding and supporting higher resolution. AVIF images have a max size of 8192x4352 last I checked. That's not a problem for web images almost all of the time, but there's certainly uses where you might want more. I think AFIV might be limited to 12bit overall while JPEG XL can do 32bit. Again, is that something many will take advantage of on the web? No, but if it's something you care about it might be really meaningful to you.
Also, if you care about the 90% quality level, JPEG XL often outshines AVIF, but when I'm putting images on the web it's often well below that.
Some of it was bad information. I saw some people saying JPEG XL being way less computationally intensive (even less than WebP) and offering much better compression. Some of it was selective data - do you only test things at 95% quality level and declare a winner?
But I think the big thing is that a lot of people wanted the two formats to succeed or fail based on their merits. I think a lot of people don't like AVIF because it isn't well supported outside of browsers - and there can be a hope that the JPEG name would mean JPEG XL would get the kind of adoption JPEG did.
And it's hard to truly compare formats - or benchmark in general. As the article shows, different JPEG XL encoders offer vastly different performance. And when we're dealing with lossy compression, sometimes it can be a bit of personal preference. At some point, people started enjoying the way that MP3 garbled symbols (the instrument) and that was the sound they expected a symbol to make.
Likewise, you look at Artificial Analysis' general index and it cost $8,708 to run with Opus and $2,407 to run with Argon so even if Argon's token price doubles, it won't be the same price as Opus.
Sonnet is cheaper than Opus per-token, but in their coding tests Sonnet burns through enough tokens that it's 7.5% more expensive.
So much depends on how token efficient something is and whether it gets you a good result or not, right?
Plus, if one is using a plan instead of paying token rates, what does my $200 get me from OpenAI, Anthropic, or Google?
It just feels very hard to compare pricing given that the number of tokens used can vary so much. In Artificial Analysis' coding tests, Sonnet used 7.1x more tokens than Sol. Sonnet and Sol might have the same per-token pricing, but there's a huge difference in the number of tokens used.
The move to the A19 also meant that they couldn't go with LPDDR4X RAM and instead had to use more expensive LPDDR5X (IIRC).
It is crazy. I expect the new Apple TV to be a flop because $300 is just a lot. The problem is that Apple is a little boxed in on this device. Apple wants to re-use the A19, but that locks them into expensive DDR5 RAM. They want to push Siri AI, but that locks them into 8GB RAM. Unlike the iPhone, there aren't as many other opportunities with the App Store, Apple Pay, search engine placement revenue, etc.
Storage has also gone up a lot and Apple is no longer selling a 64GB model. Most streaming sticks have 4-16GB of storage and it's usually really slow/cheap storage. The new Apple TV has 128GB of storage and the price for that has probably gone from around $25 to $50.
You say that it's fine from "$200 to $300", but that's not really true. It's gone from $250 to $300 when you look at the 128GB model. I still think it's crazy that it went from $150 to $300 in the same year, but the outgoing 128GB model was already $250 after the price hikes earlier this year.
The problem is that Apple didn't want to reign in their product for the current market and I somewhat get that. What's the point of a new Apple TV if it's just "we updated the processor." For me, the point would be AV1 and VP9 decoding, but Apple wanted to push Siri AI.
Other high-end TV devices have also seen big increases in price. The Fire TV Cube went from being $110-140 to now $200 - and with the Fire TV Cube you're getting Cortex-A73 processor cores from 2016, 2GB LPDDR4X, and 16GB storage. At least Apple is giving you 8GB LPDDR5X and 128GB of storage and a processor that is way better.
The Google TV Streamer is $150 now with 4GB RAM and 16GB storage, but with LPDDR4 RAM and Cortex-A55 cores - and it's still $150.
A Roku Ultra is 2GB RAM and 4GB storage with Cortex-A55 cores and costs $150.
I think we're seeing that there's still some really old stock out there that's cheap, but we're seeing the prices climb on competitors too. Google raised its price 50%. Amazon has raised its price 40-80%. Apple raised its price 54-67% on the same-model and launched the newer model with double the RAM, more expensive RAM, a processor that's 4 years newer, etc.
I think it's likely that Google, Roku, and Amazon simply don't upgrade their streaming devices during this parts shortage. I think a lot of people will be fine with a 12nm Cortex-A55 based TV streamer with limited RAM and storage (and slow RAM and storage). I love how fast my Apple TV is. It's amazing the difference in UI lag. But I think most companies aren't going to prioritize that premium experience, especially during a parts shortage.
People care that macOS has a lot of compatibility with Linux/BSD, but most don't care that it's UNIX® - just as people don't care that Linux/BSD aren't UNIX® compatible. Yes, Linux/BSD are very similar to UNIX® and mostly compatible, but they don't conform to the specification.
You (probably) don't care about the UNIX® specification. You care that your MacBook has good compatibility with Linux/BSD infrastructure. Many people assume that Linux/BSD are UNIX® and just don't pay for certification. They aren't UNIX® compatible. They are extremely similar, but they don't follow the spec 100%.
Sorry for putting the ® everywhere, but people often use UNIX when they mean nix-like operating systems and when we're talking about whether people care about macOS as a UNIX® or not, it matters whether were saying "people want macOS to be UNIX®" vs "people want macOS to have good compatibility with nix-like operating systems such as Linux and BSD like it does today." People usually care a lot about the latter and not about the former.
> Is this Apple just not wanting to pay for the certification anymore or does it actually change anything for us devs?
It's probably not about paying and more about tiny differences. Will it change anything for you? Almost certainly not. In fact, it might make things better for you. If anything, Apple's adherence to the UNIX spec can sometimes mean subtle differences in tools between macOS and BSD/Linux. Do you want things like grep or sed that conform to the UNIX spec or that work like grep and sed work in the real world?
And these differences are often small edge cases that don't come up much, but a spec is a spec.
The money is inconsequential for Apple, but at this point it probably makes more sense to just stop calling it UNIX and not worry about shipping a version of a tool exactly meets the UNIX spec - it'd be more useful for their users if it matched the behavior on Linux/BSD.
I think the issue is time and money. Toyota maybe doesn't want to spend the money on a completely different car in the compact sedan category just for EVs - at least not yet. Plus, it might mean waiting several years when they want something out now. Sometimes it's about tradeoffs like time and money. Adapting an ICE Corolla might be reasonably cheap and fast. They might feel the need to have an EV Corolla out sooner than later. They might want to take their time with a new platform EV so that they don't put out something that doesn't meet their quality standards. While the drivetrain might be different, everything else in the Corolla is battle tested.
I think a lot of people are attributing speeds of 200-700Mbps to mm wave when it's actually pretty ordinary speeds for 2.5-5GHz coverage today now that carriers have decent capacity there. If you're getting below 144Mbps, you're in the bottom 25% of T-Mobile's network experience. Over the past 6 years, speeds have gotten a ton faster, mostly from 2.5-5GHz spectrum. Verizon is the only carrier that put much effort into mm wave and it still covers basically nothing.
Ignoring latency, you're likely to get speeds similar to wifi from today's mobile networks without mm wave - unless you have a really great home internet connection and a good wifi setup. I have a MacBook Pro, but it only has wifi 6 and I can get 700-800Mbps out of my wifi, but I'm not getting 2Gbps from it.
C-Band spectrum also makes sense given a speed test of 400-800Mbps. That's fast, but not hugely faster than what you should expect with really good coverage today. T-Mobile experiences speeds above 560Mbps 25% of the time. Verizon's 5GUW (which is mostly C-Band) experiences speeds above 620Mbps 25% of the time. You ran two speed tests and got 477Mbps and 767Mbps. One of those is pretty average for C-Band and one of those is quite good for C-Band (25% faster than the 75th percentile), but nothing too out of the ordinary.
Those are speeds that are relatively common when carriers have around 150MHz of 2.5-5GHz spectrum deployed in your area and you have good signal and not too much congestion.
mm wave does exist, but the coverage is often a block or two with indoor coverage being really spotty. C-band coverage will be measured in square miles. C-Band won't have the coverage that lower frequencies have, but it's a lot more practical to deploy than mm wave. mm wave shines when you have a small, high-value area where you want to deploy 1,000MHz+ of capacity. Carriers might have 50MHz of capacity around 800MHz, 90MHz of capacity around 1.9GHz, 150MHz of capacity around 4GHz, and 1,000MHz of capacity around 30GHz. That really high frequency mm wave won't cover a large area, but there's a lot of it. If you're looking to cover a sports stadium, it makes sense. If you're looking to cover people's homes (even in major cities), C-Band is going to get you really good speeds with coverage that will be a lot more usable.
It's possible that you're getting mm wave, but those speeds are pretty normal for good C-Band coverage. The first test is pretty average and the second test is quite good, but nothing extraordinary.
There's a lot of tires that are poorly made and just bad tires. For example, one tire got 2/5 for rolling resistance, tested to 55,000 miles for tread, and 2/5 and 3/5 ratings for the traction categories like wet braking, snow traction, ice braking.
I'm not arguing that there aren't trade-offs, but tire companies have been working to make better tires that minimize the trade-offs and produce better tires on all three.
> The rule is “designed to ensure that replacement tires sold in the state are at least as energy efficient, on average, as tires sold in the state as original equipment.”
It sounds like the rule isn't exactly that strict. This (combined with the fact that all season tires aren't covered) presents a very low bar. But there are tires that are just garbage - they're bad at all three. It doesn't sound like this rule would impact good tires that have very long tread life and amazing traction. It sounds like it'd impact the tires that are a pile of garbage.
There's a difference between a tire that trades off a small portion of rolling resistance for traction and tire wear and tires that are poorly made and are trading off a huge amount of one or two of those categories because they're just bad tires.
Per-token cost isn't a great metric given that some use way more tokens than others.
They end up wasting their energy on work that won't help them reproduce and needing to waste more energy getting rid of it.
For example, if a data center is illegally running gas turbines, the penalty could be that the data center must be shut down for 3 years and no equipment can be removed from the data center during that time.
With financial penalties, there's always calculus on whether the penalty is high enough to deter the action. But the only reason they'd want to, for example, use illegal power generators at a data center is to run the data center. Losing access to 100% of the value of that data center for 3 years would make the calculus easy: it wouldn't be worth it for the companies. They'd lose the entire reason for doing it and so much of the capital investment in equipment as it depreciated during those 3 years.
With AI companies believing the market is worth trillions, it's hard to imagine a financial penalty high enough to deter illegal actions. It might need to be in the hundreds of billions - a level which they probably can't pay given their current resources and which governments wouldn't levy. Not that they'll essentially confiscate a data center and its hardware for multiple years as a penalty either.
There are parts of states like Grant County Washington that have cheap hydro power, but it's very rare for power to be that cheap in the US. Even if this applies to you, it won't apply to the vast majority of people on here who will have electric rates 2-4x higher.
Average electric rates by region:
New England 28.1 cents
Mid Atlantic 25.1 cents
East North Central 20.8 cents
West North Central 14.8 cents
South Atlantic 16.1 cents
East South Central 15.5 cents
Mountain 14.6 cents
Pacific Contiguous 26.1 cents
Pacific Noncontiguous 42.1 cents
Models that people like the design aesthetic of (Claude, GLM) tend to do better in LMArena than they do on other benchmarks. Design matters, but you look at a model like GPT-5.5 and it's behind Kimi K2.6, Sonnet 4.6, Qwen3.7 Max, and GLM-5.1 on LMArena's code leaderboard. Then you look at benchmarks like DeepSWE and GPT-5.5 blows them out of the water with only Fable and GPT-5.6 beating it.
I'm not saying that the LMArena leaderboard isn't useful, but I'm not sure how much weight I'd give it as a "code" leaderboard. I think often times it's a design comparison of simple front-end React apps rather than a coding comparison. GLM-5.2 is a very good model, but when you look at DeepSWE or Terminal-Bench v2, GPT-5.5 is well ahead.
At this point, I always look at things like Artificial Analysis' total cost to run their tests. It'll take into consideration the cost of tokens, how many tokens it burns through, and how effectively it uses caching (and the price of that caching).
If a model "costs the same" but its reasoning ends up going through a ton more tokens, it doesn't really cost the same in real world usage.
I feel like the "something new" was price. Over time, that price kept creeping up. Yes, it went from being a 95% flagship to being a 100% flagship, but it also went from being half price to full price.
It was also cool that it used Cyanogenmod which meant you got a community OS that actually got updates, but over time other manufacturers started offering updates for their phones (rather than abandoning them soon after manufacturing). And that was something new other than price. But I think the big thing was that it was a half-price phone when it launched. In 2014, it was just such an amazing deal. Today, it's the same price as Samsung phones.
- Telegram had usernames in 2014 before Signal added them a decade later, allowing people to chat without sharing their phone number
- Telegram has unencrypted chats which allow for giant chat rooms of 200,000+ and channels with millions of subscribers. Signal warns about performance issues when you have more than 150 people in a group. Telegram isn't just a messenger - it's often used as a social publishing platform like Instagram.
I don't use Telegram and use Signal a lot, but I also understand why other people use Telegram: the same reason they use Instagram.
Cursor's benchmark finds that Cursor's model (Composer 2.5) is basically as good as Opus 4.8 max and GPT-5.5 xhigh, but at a fraction of the price.
Artificial Analysis' testing shows Composer 2.5 to be pretty far behind: https://artificialanalysis.ai/agents/coding-agents. You look at the DeepSWE benchmark (which is probably the hardest to game at this point) and GPT-5.5 xhigh gets a 64, Opus 4.8 max gets 56, and Cursor 2.5 gets 16.
I don't doubt that Cursor works well for some people. It's beating DeepSeek v4 Pro in the DeepSWE benchmark and that's a very capable model. But I'm skeptical of the claims that it's a competitor for Opus 4.8 and GPT-5.5. It just seems convenient that their model does so well on their own benchmark while third party benchmarks have it far behind. Maybe it's a really great benchmark and a better measure than third party ones - I'd love for a cheap model to do as well as the expensive ones.
It's not simply marketing since the Pro/Max chips of a generation use the same cores as the regular version, just more of them or different combinations of performance and efficiency cores.
They hold the keys for your DID. If they don't allow you to move to another PDS, you can't move. The original theory was that you'd hold the private keys, but that's something that would hugely limit adoption so they decided to hold the keys themselves.
In terms of moving your backlog of posts to a new server, part of the issue is liability (not merely legal liability, but reputational as well). When you have a user on your platform and they're posting stuff, you're moderating them in real time. If they turn out to be a horrible troll, you've get the reports. Let's say a horrible troll has been on EvilServer and EvilServer has been ignoring the reports against them. They now want to move to your GoodServer and bring all their post history with them. As an admin of GoodServer, you can't see that everyone has been reporting this troll for years. They're now moving over lots of horrible, inflammatory, potentially illegal posts to your server.
In the coding index, GPT-5.5 gets 59.1, 58.5, 56.2, and 52.1 for xhigh, high, medium, and low while Muse Spark is behind at 47.5. For agentic, GPT-5.5 gets 74.1, 72.0, 69.4, and 59.7 (xhigh, high, medium, low) while Muse Spark gets 62.0 (beating only GPT-5.5 low).
GPT-5.5 only gets beaten by Opus 4.8 in their general index, is the top spot for coding, and is #3 behind Opus 4.8 and GLM-5.2 for agentic (excluding Fable 5 which takes the top spot, but is unavailable).
For me, the big thing with MiMo-V2.5-Pro and DeepSeek V4-Pro is that cached inputs are practically free. Kimi K2.7 Code is 53x more expensive for cached inputs which is 95% of my costs.
If I use 95M cached input tokens, 4M input tokens, and 1M output tokens, that'd be: $18 for cached input on Kimi K2.7 Code vs $0.34 with MiMo/DS; $3.80 for inputs on Kimi vs $1.74 with MiMo/DS; and $4 for output on Kimi vs $0.87 with MiMo/DS.
Of all the places where I'm accumulating costs by using Kimi, it's the cached inputs. The real savings with MiMo/DS's price cut is the cached inputs.
Qwen-3.6-27b is closer to Claude Opus 4.7 than it is to Haiku 4.5 in a lot of benchmarks - and it's way smaller than Microsoft's new model.
Sure, it competes with Haiku, but it shows how far Microsoft is behind lots of other small models that are available.