Yup, while I'm sure that part of it is that the big players got used to this modus operandi, part of it it just your client not knowing what they want. I work for a financial institution, I see this everyday. Our IT partners just have to wait for our ill-conceived and ill-defined tickets and changes of priorities, keep track of the tickets and the money keeps flowing by itself. I know the analogy isn't perfect, but still.
Hats off to Anduril for shaking up the space. Wonder if the Swedes will be able to match them.
The client not knowing what it wants, and then having multiple clients who all want different things. Making meaningfully different variants for different branches of the US military seems excessive.
Do we really need a VTOL version? I don’t know, maybe. Or maybe it looks cool and we can do without / spend that money elsewhere.
We don't need it, but it's too late now to drop it. Going back the the Joint Strike Fighter program circa 1993, the primary requirement for the V/STOL "B" model came from the USMC. They had an institutional fear stemming from experience in WWII that they couldn't depend on Navy fleet carriers for the close air support mission so they insisted on having that capability organic to amphibious ships (which lack catapults and arrested landing equipment). Originally that mission was performed by the AV-8B Harrier, but they (correctly) recognized that it wasn't survivable against modern air defenses. Hence the requirement for a 5th-generation replacement. The Marines had a lot of political influence in Congress at the time and were able to get their way. Requirements from international program partners including the UK and Italy who wanted to pretend that their crappy little mini-carriers could still be relevant also played a part.
Today the notion of conducting an opposed amphibious assault looks increasingly unrealistic. Even third-world terrorist groups now have drones and guided missiles that could massacre the connectors and transport rotorcraft. The F-35B can survive but there won't be any ground troops alive to support. So it's original reason to exist has largely vanished. And the sad thing is that the pointless V/STOL requirement forced a fat single-engine airframe on the entire program which badly compromised the "A" and "C" models in terms of reliability, range, speed, and payload. Oh well.
The whole thing of nations trying to build the fastest, stealthiest, VTOL-est, most high-tech expensive fighter they can reminds me of the generals desperately trying to engineer cavalry charges in 1917 or the navy folks dreaming about bigger battleships in 1944. To see where this ends up, look at Ukraine, you've got a few of the remaining high-tech super-expensive fighters loitering a long way from the battlefield launching long-range standoff munitions and then leaving again, and that's it. Or look at Iran, lawnmower-engine-powered darts stalemating the country with the most expensive, high-tech fighters in the world, and the US is running out of munitions while Iran is a long way from running out of darts. A thousand darts beat one F-35 no matter how technically superior it is.
But I always preferred the F22. That thing is beautiful and the metrics are insane. But for a while it seemed like there was no peer contender and they stopped the programme.
The F-22 cannot do half the things the F-35 can. Yeah it's fantastic as an air superiority fighter, but the multirole versatility makes the F-35 the one plane most air forces need to have.
Huh, never thought about that 2nd point - I'm also transitioning to pi, but didn't think to change to neovim as well. I guess your logic holds for claude code users as well though.
But isn't it the case that we can't reach this safeguard with the current architecture? I remember Karpathy making an interesting point 2 years (cca) back, that I would summarize somehow like this: the mechanics behind every LLM answer are the same, what you then call hallucination is more or less a consequence of whether or not the answer was factually correct/useful.
Which would mean, as is so often the case, that the "killer feature" of the LLMs is also its biggest weakness and the two can't be disentangled. Now, we are inventive creatures and we might come up with a remedy for these issues, but what you basically see so far is more guardrails, the use of harnesses and building a whole bunch of infrastructure around the LLMs to get useful work out of them.
Which, btw is not a critique, I do it as well and it's a fun engineering challenge.
Maybe less shocking, but still, so is google. I know they are one of the labs, but that being said they rely far less on gemini for revenue as opposed to the big 2.
I think Google can play both sides here: they may end up a frontier lab if OpenAI/Anthropic dissolve, or a regular LLM provider, in which case they could be serving open models.
It seems like they're sticking to a static pricing plan that was made when the only relevant competition were the US labs (im not counting deepseek 2025 as serious competition -> glm and then kimi on the other hand, now that's a different story).
This was exactly my thought process. Coupled with a less idealistic one: mid 2025 if you looked arena ELO scores google seemed to be dominating and I was sure the trend would continue. But they really dropped the ball on coding and tool use.
That being said, controlling android and apple mobile devices is kind of a big deal. And their video models are still top notch.
Exactly my experience. I'm building an AI document-extraction platform, so I had to benchmark a bunch of models — on the cost:quality:latency:adherence picture, flash wins hands down for structured extraction. (Caveat: I've only tested the three US labs and Mistral.). So like u said, totally viable in prod for a relatively static tool. Didn't build the tool suite with gemini, but if you use service mode it currently mainly runs on flash.
Haven't done any serious coding work with the flash models though — but I'm seeing more and more HN comments from people who seem to have picked it up for that in the last couple of months.
Haven't been following the articles and snippets we get from these labs about training their models for a while. But I'm guessing the latest chinese models are way less based on distilling? If not, then your speed of progress is still limited by the two labs (which we are collectively, in various forms subsidizing).
Hats off to Anduril for shaking up the space. Wonder if the Swedes will be able to match them.