Frank Angelo Albanese (@faalbane) header
Frank Angelo Albanese (@faalbane) profile picture

🙏📍 NYC · YC W21 · Antler W23 · Independent AI Researcher

Joined February 2016
Photo by Frank Angelo Albanese (@faalbane)Photo by Frank Angelo Albanese (@faalbane)Photo by Frank Angelo Albanese (@faalbane)Photo by Frank Angelo Albanese (@faalbane)Photo by Frank Angelo Albanese (@faalbane)Photo by Frank Angelo Albanese (@faalbane)Photo by Frank Angelo Albanese (@faalbane)Photo by Frank Angelo Albanese (@faalbane)Photo by Frank Angelo Albanese (@faalbane)Photo by Frank Angelo Albanese (@faalbane)
The classic case against releasing weights: once they're out, nobody can monitor how they're used, revoke access, or change the model's behavior. Read that list again as a buyer. No one watching your prompts. No one able to cut you off. No one changing your model under you overnight. For a company putting its IP into a model, those are the features.
13
Microsoft at Ignite, launching weights customers can adapt inside their own Azure: "More of what makes a company distinctive will live inside the AI models it runs." Their example: Alaska Airlines tuned a model on aircraft maintenance reports and took classification accuracy from 81% to 91%. When the biggest vendors start selling control over the weights, the question isn't whether companies will own their models. It's what they'll teach them.
22
Tencent put Hy4 Flash's weights on Hugging Face, and by this morning other people had already published GGUF, MLX, NVFP4, AWQ and FP8 builds of it. That's the part an API can't match. Ship weights and within a day the community has them running on Macs, consumer GPUs and whatever hardware you already own. Ship an endpoint and you get one vendor's roadmap.
14
"Open weights" describes a download, not a license. MiniMax's new H3-Image license requires written permission from MiniMax once you pass $1M in annual revenue or offer it as a hosted service, and it bars using the model's outputs to improve other models. Before you build on an open model, read the LICENSE file, not the headline. Apache 2.0 or MIT means you own your roadmap. A revenue gate means someone else still does.
91
Baidu shipped ERNIE 5.5 Flash Base on Nov 5 and a post-trained Preview on Nov 11, which it calls an intermediate checkpoint that improves as post-training continues. Same pretrained base. New product. That layer is where the leverage is, and you can run it on your own data.
45
"Open weights need a data center" was fair. Edge0's 145B model activates ~10B params per token, and its model card shows it on one 24 GB RTX 5090 D (23 GiB VRAM plus 38 GiB system RAM). The hardware excuse for not owning your model is expiring faster than procurement cycles.
1
38
Linux v. Windows is a helpful example from history to illustrate how open source has historically challenged closed source. The difference with AI, is that the barrier to entry is virtually eliminated, which makes open source AI the de facto global winner of future market share.
28
Frank Angelo Albanese retweeted
open models are how longevity escapes the big companies. when anyone can run a biology model on their own machine, the cure doesnt need anyone's permission. no board can fire that
8
11
1
154
4,854
The most ironic outcome in AI is that the most profit-hungry private labs will be the reason why open source prevails to the point of RSI and beyond.
15
It's funny and ironic that people are contributing to Nvidia's Molt repository with Claude, arguably one of the most effective toolkits for post-training open-weights models, which is arguably the biggest threat against Anthropic's market position. github.com/NVIDIA-NeMo/labs-…
34
Anthropic's current market position will be seen as a flash in the pan of AI history. It's only a matter of time before every company on Earth has post-trained private models internally, and then it's the dot com boom all over again.
2
2
25
The irony is that people will use frontier closed AI to develop better open-source models, and then dump their subscriptions.
11
If self-improving AI is inevitable, then open-source is inevitable. If open-source is inevitable, then 'large AI company moats' are inevitably dispensible.
1
15
Wow, Microsoft Outlook appears to be inaccessible.
2
4
1,283
Oh the amount of times WhisperFlow has autocorrected 'Grok Bot' to 'Crock Pot'. Now that I think about it though, 'Crock Pot' would be a funny name for an open-source Grok Bot.
3
6
420
Love is the only inexhaustible resource.
1
2
139
I’m far more terrified of life without love than death without life.
2
125
One of my favorite (and seemingly underrated) uses of LLMs is clearing up memory on disk. Codex Sol 5.6 is particularly useful for this. 'Freed another ~18 GB.'
2
95
In the best of us, the greatest good is greater than the worst evil in the worst of us.
1
3
90
The age of superintelligence is like extreme versions of the best and worst of humanity projected at the same time. There will be good and evil, just like there is now. The important thing is to balance these as superintelligence grows.
2
2
77