Meta Delays Rollout of New AI Model After Performance Concerns 27
Meta has delayed the release of its next major AI model after internal tests showed it lagging behind competing systems from Google, OpenAI, and Anthropic. The New York Times reports: The model, code-named Avocado, outperformed Meta's previous A.I. model and did better than Google's Gemini 2.5 model from March, two of the people said. But it has not performed as strongly as Gemini 3.0 from November, they said. As a result, Meta has delayed Avocado's release to at least May from this month, the people said. They added that the leaders of Meta's A.I. division had instead discussed temporarily licensing Gemini to power the company's A.I. products, though no decisions have been reached.
[...] It takes time to improve A.I. models, and Meta can still catch up to rivals, A.I. experts said. But a longer timeline has set in at the company, with Mr. Zuckerberg tempering expectations for Avocado in the past few months. "I expect our first models will be good, but more importantly will show the rapid trajectory we're on," he said on a call with investors in January. A Meta spokesperson said in a statement: "As we've said publicly, our next model will be good but, more importantly, show the rapid trajectory we're on, and then we'll steadily push the frontier over the course of the year as we continue to release new models. We're excited for people to see what we've been cooking very soon."
[...] It takes time to improve A.I. models, and Meta can still catch up to rivals, A.I. experts said. But a longer timeline has set in at the company, with Mr. Zuckerberg tempering expectations for Avocado in the past few months. "I expect our first models will be good, but more importantly will show the rapid trajectory we're on," he said on a call with investors in January. A Meta spokesperson said in a statement: "As we've said publicly, our next model will be good but, more importantly, show the rapid trajectory we're on, and then we'll steadily push the frontier over the course of the year as we continue to release new models. We're excited for people to see what we've been cooking very soon."
Obligatory (Score:2)
"That's what she said."
Re: (Score:2)
Leapfrog (Score:2)
For a company to show AI Leadership (whatever that means), the new model must leapfrog the previous released model numbers (from your competitor). Mark does not want to appear to be behind Google's previous model (it would impact his stature).
I wonder which company will figure out how to successfully game the numbers first.
Re: (Score:2)
our next model will be good but, more importantly, show the rapid trajectory we're on,
That statement shows that their next model will not, in fact, be good; otherwise they would focus on the goodness not the trajectory.
Re: (Score:2)
Come on!
They're converging already!
They're converging! They're converging!
They're converging! They're converging! They're converging! They're converging!
Oooooh.
I Feel Conflicted (Score:2)
On the one hand, Zuck is pure evil, and I hate all things Meta.
On the other hand, their model is one of the best open models, and I want open models to succeed.
Not sure whether to cheer this news or be disappointed.
Re: (Score:3)
Is the training data open? No? Not an open model then.
Re: I Feel Conflicted (Score:2)
It's open weights, not open source. But the word "open" meant documented and interoperable before it popularly was applied to source code, so yes, that's still a kind of openness. Open systems, open standards...
Re: (Score:2)
Open Weights is more imporant, no matter if the OSI would like to have the training data. The training data is only useful if you have 5 million to train the model again (and why would you want to, if you in the best case obtain the same weights?) whereas open weights allows you to train much cheaper the things you need on top of a model that has already sufficient pretraining.
Re: (Score:2)
I haven't formed an opinion on which is more important, but it's clear that open weights are more immediately useful.
Re: (Score:2)
I don't think anyone would say no to get also the data made available, but most won't even download it. Such data is at the scale of hundreds of Terabyte for image models and I don't think much less for text models (which are in average larger than image models, one should not underestimate it just because it only text).
Re: (Score:2)
I'm thinking more along the lines of distributed training. Nobody would need the entire corpus.
Re: (Score:2)
I think there are currently no good approaches to that. Each iteration needs the back propagated error of the one before. Splitting up a single pass is probably too inefficient, and distributing gradients (e.g. to accumulate from different participants) also needs low latency. Most neural networks are already limited by RAM bandwidth (in particular LLM), and that's quite a few orders of magnitude faster than network IO.
Re: (Score:2)
It also means you cannot get training defects out.
Re: (Score:2)
Llama 3 was the last good Meta model. Llama 4 basically failed. After that there came are lot of better models than both, like Mistral Large 2 (Coder, Ministral variants), Qwen3 and later Qwen 3.5 and many more. The next good US model to be expected is probably Gemma 4.
The AI IQ asymptote is coming (Score:3)
Let's hope their bankruptcies don't have too much effect on real businesses built on real numbers.
And of course: GPU, HDD, SSD and RAM fire sales !
Re: (Score:3)
Re: (Score:2)
Fire sales on server equipment that almost no one can use in their home computer.
The OCP (Open Compute Project) solutions are optimized for the hyperscaler datacenters, and reuse in a home lab after retirement is not a design goal (well, unless you have a datacenter in your home). Given the demand for AI hardware (and profit to be made) the major vendors are likely prioritizing solutions that are OCP (or other hyperscaler designs) compliant, which will dry up the used market at the lower end that was a sweet spot for the home lab users.
Re: (Score:2)
If the fire sale discount is deep enough I would love an AI server for my home running a local open source model lol.
Re: (Score:2)
Re: (Score:2)
Even if that intelligence grows linearly with time, they will soon reach super-human level and, it is assumed, AGI.
But what will actually happen is that the LLMs' actual, usable intelligence will level out with time and will asymptotically approach a maximum level for this technology. They have run out of text to input and adding more and more compute ( using our energy ) to add more and more parameters w
Re: (Score:2)
But it won't be that cunt Altman who does it.
LOL tell me how you really feel about Altman, don't hold back.
Re: (Score:2)
An ongoing rewrite (Score:2)
From what I've heard the incredibly well compensated engineers Zuck poached from the competition have had to rewrite much of Meta's AI architecture. Presumably Meta had retained all of their corpus of training data for loading into the new model but they are chasing a rapidly moving target. Google's Gemini 2.5 is ancient history (meaning several months old). Gemini 3.1 is widely available but it isn't nearly as good at coding as the most recent Anthropic models. I'm skeptical that Meta can ever catch up.
Re: (Score:2)
Re: (Score:2)
>> Why didn't they use AI to rewrite it?
They did of course. But it's too late to catch up. Let's say you have shamefully migrated from a top tier AI company to Meta in exchange for a gigantic amount of cash. You are intimately familiar with large swathes of the former employer's technology and may even have invented some of it, but you must start over from scratch.
Even with AI assistance are you going to reach parity with the firmly established leaders in a couple of months? That's the timeframe for s