This likely isn't AA putting its finger on the scale for specific interests, rather in the case of Muse/Astra, it's to avoid looking really bad because it scores things way out of wack with people's anecdotal vibes on model capabilities. It is possible that labs are getting preliminary AA results back and raising concerns about scoring that then get folded into updated benchmarks, but I strongly doubt it's nefarious.
Likely they did a distillation run on a later checkpoint of a large internal teacher model, probably with some tuning of the distillation process based on early public results with Argon. RL is the expensive/time consuming step, they should be able to distill down from a checkpoint with the stamp of approval pretty quickly.
Even a decade ago marketing/growth was the massive lion's share. IMO just building something has been near meaningless since at least post-Ruby on Rails. The app stores have been inundated since the late 2010s.
The distinction is adjusted vs non-adjusted basis. Serving inference is profitable, but serving inference for the last released model doesn't cover the cost of the model in training.
Dario explained it very clearly in an older interview. If every model is bigger and more expensive to train, each model can individually be profitable, while the company as a whole still shows a loss at any given point in time. This is why he likes the idea of regulation so much, it'll let them take a breather and recoup.
Anthropic cooked with Opus 4.5 and 5.5. Their other iterations have usually been nicer to interact with than GPT and more "autonomous", while being dumber.
Agents are already better at writing strongly typed languages than Ruby, and don't care about the small ergonomic affordances that rails provided. Even if you think the Rails bloat and Ruby nicheness/slowness is worth it today (hard disagree), think forward a year when agents are significantly stronger than they are today and consider how that decision will hold up.
IMO Cloudflare is the closest thing to a "good guy" these days. Very customer friendly, affordable and generally fairly ethical (if you pass over the part where they share data with the NSA).
It's insane to me that people don't see that agents are going to be how you interact with your computer going forward, almost exclusively. The people who make things hard for agents are going to get eviscerated by the people who don't just make it easy for agents to do things, but market primarily to agents rather than humans.
As a Krea2 local model enjoyer, your challenge is that on the $20 GPT plan you get a fairly good amount of Codex use and image gen. If I used Krea, it would be because I fine tuned/built a workflow on top of it to do something GPT image couldn't/wouldn't do. Maybe there's a market for the cheapest "good enough" bulk image gen model (spam? idk what the applications would be), but personally I'd say you should focus on workflow add-ons/one click fine tuning/etc that design companies can use to create assets with their custom "look" since the GPT-Image look will become blase from overuse.
The delta isn't really slower, we're just unlocking capability bundles in batches. The capabilities bundles you care about might have mostly been unlocked already, and just need refinement, which is coming slower than the initial capabilities came, but don't discount the breadth of advances being made across all possible tasks.
Beyond SVG, these things have been RL'd on 3d reconstruction (which carries over), and I expect given the public reaction to recent frontier advancements in 3d capabilities that this will get allocated a larger share of the post training budget in the future.
Counterpoint, lots of people learn to navigate the areas where they spend their time by following GPS directions repeatedly, and can drive regular routes without them (though they're still useful for routing around unknown traffic issues).
The ironic thing is the head chefs of restaurants do almost no cooking. They order around the sous chef, who orders around the kitchen staff, while being focused on the big picture.
First, just check out the comments on any "I'm tired of reading slop" submission to refute your first point, the circle jerk is real.
Second, while I sympathize with not wanting to be inundated with low effort PRs, putting a ban on AI assisted contributions in response is myopic virtue signalling.
I don't choose TS/JS for new projects anymore unless they're browser based. If you need to be in JS land for whatever reason, deploy with Node but do local package management/testing with Bun, because it's way faster for those and you can migrate off without too much pain if it ever becomes problematic.
That benchmark agrees with my point, GPT-4o scored an 80% on it. Even if you could construct a much harder legal bench that could resolve smarter models, the bottleneck is getting the absolute top legal experts in the world to give feedback, so progress would be slow compared to stuff like math and coding where you can programmatically generate scenarios and validate/score performance.
These models are spikey as hell, they can gain incredible capabilities but that doesn't make them godlike minds, it's more like a scaled up version of rain man.
It's hard to say for sure, I think we can at least confidently assert that games are going to be more challenging than other forms of creative expression like Books, Music, Art, etc because people's tolerance for derivative games is lower than their tolerance for derivative movies/music.
Agents' ability to disassemble and reverse engineer is amazing, the scary thing to think about is what does it mean to sell software in a world where the only moat is brand/marketing?
I don't think AI will ever really solve "making a fun game" because:
1. Games are highly complex systems, so for example adding a fun mechanic from one game to another game can actually make that game less fun, and AI has no way to know, you just have to playtest it.
2. Games are evergreen (people still play and love the original Zelda/Mario/Tetris) and social (you want to share your love of game XYZ with others), so there's literally no reason for anyone to play a purely derivative game, when you can just play the original.
Eventually AI will be able to make perfect assets and all the other stuff AROUND making a fun game, but the game design will always be a human thing. At best, AI will come up with a bunch of game design experiments that a human will need to playtest and provide feedback on.
In terms of token and attention efficiency, lisp's syntax is actually not quite optimal. Ironically, something like xml where you can munge the delimiters together with the tag as a single token for common cases (e.g. '<tr>' and '</tr>' vs '(' 'xyz' ')') is better.
A little nit here, bash isn't the most elegant language, but shell languages have a very nice feature that makes them easier for Agents to write: they're executed left to right, in the same order the tokens are generated. Python/Javascript can execute out of token order (e.g. foo(bar(baz(bak(bat))))), which is more error prone. Additionally, shell is very terse and designed for duct-taping things together.
I think something like nushell is probably a better direction than trying to use python/javascript as a shell.
My argument is that if you were to compare "analyze XYZ geopolitical situation" or "explain the ramifications of XYZ law" from Opus 3.5, 4.5 and 5.5, the difference would be marginal, at least for 4.5 - 5.5. Almost all the crazy capabilities newer models have is from RLVR variants, whereas capabilities driven by RLHF are inching along.
The magnitude of improvement in unverifiable domains is small, mostly down to models doing more careful research before answering and hallucinating less. They are more thoughtful, but I expect you'd have to drop 2 major versions of Opus before you'd start to see most people really clearly be able to differentiate them.