"What's some frontier research Meta has shared in the last couple years?"

the current Meta outlook is embarassing tbh, the fact they have largest data of social media in planet and they cant even produce a decent model is quiet "scary" position

Yann was a researcher not a productization expert. His departure signals the end of Meta being open about their work and the start of more commercial focus.

Llama 4 wasn't great, but Llama 3 was.

Do we all forget how bad GPT 4.5 was?

OpenAI got out of that mess with some miraculous post-training efforts on their older GPT-4o model.

But in a different timeline we are all talking about how great Llama 4.5 is and how OpenAI needs to recover from the GPT 4.5 debacle.

Just because they are not leading current sprint of maximizing transformers doesn't mean they're not doing anything.

It's not impossible that they asses it as local maximum / dead end and are evaluating/training something completely different - and if it'll work, it'll work big time.

Just because they have that doesn't mean they're going to use it for training.

I’ve long predicted that this game is going to be won with product design rather than having the winning model; we now seem to be hitting the phase of “[new tech] mania” where we remember that companies have to make things that people want to pay more money for than it costs to make them. I remember (maybe in the mid aughts) when people were thinking Google might not ever be able to convert their enthusiasm into profitability…then they figured out what people actually wanted to buy, and focused on that obsessively as a product. Failing to do that will lead to failure go for the companies like open AI.

Sinking a bazillion dollars into models alone doesn’t get you shit except a gold star for being the valley’s biggest smartypants, because in the product world, model improvements only significantly improve all-purpose chatbots. The whole veg-o-matic “step right up folks— it slices, it dices, it makes julienne fries!” approach to product design almost never yields something focused enough to be an automatic goto for specific tasks, or simple/reliable enough to be a general purpose tool for a whole category of tasks. Once the novelty wears off, people largely abandon it for more focused tools that more effectively solve specific problems (e.g. blender, vegetable peeler) or simpler everyday tools that you don’t have to think about as much even if they might not be the most efficient tool for half your tasks (e.g. paring knife.) Professionals might have enough need and reason to go for a really great in-between tool (e.g mandolin) but that’s a different market, and you only tend to get a limited set of prosumers outside of that. Companies more focused on specific products, like coding, will have way more longevity than companies that try to be everything to everyone.

Meta, Google, Microsoft, and even Apple have more pressure to make products that sanely fit into their existing product lines. While that seems like a handicap if you’re looking at it from the “AI company” perspective, I predict the restriction will enforce the discipline to create tools that solve specific problems for people rather than spending exorbitant sums making benchmark go up in pursuit of some nebulous information revolution.

Meta seems to have a much tougher job trying to make tools that people trust them to be good at. Most of the highest-visibility things like the AI Instagram accounts were disasters. Nobody thinks of Meta as a serious, general-purpose business ecosystem, and privacy-wise, I trust them even less than Google and Microsoft: there’s no way I’m trusting them with my work code bases. I think the smart move by Meta would be to ditch the sunk costs worries, stop burning money on this, focus on their core products (and new ones that fit their expertise) and design these LLM features in when they’ll actually be useful to users. Microsoft and Google both have existing tools that they’ve already bolstered with these features, and have a lot of room within their areas of expertise to develop more.

Who knows— I’m no expert— but I think meta would be smart to try and opt out as much as possible without making too many waves.