ML @premai_io | FPV drone Pilot | 2x @join_ef Founder Prev @thequillbot, @ibmresearch
Joined October 2015
- Tweets3.6K
- Following829
- Followers811
- Likes16.3K
Pinned Tweet
I am excited to announce that my abliterated qwen model hit 1M+ downloads on @huggingface.
The demand for unbiased and not overly censored models has exploded exponentially over the last few weeks. I am also excited to announce that @premai_io router now serves abliterated models like Kimi K3 and GLM 5.2. Do check us out.
Huggingface: huggingface.co/collections/B…
PremAI: premai.io/blog/kimi-k3-prem-…
AI companies bought up all the compute, HBM and manufacturing capacity. Now Democrats want to pause new AI data centers.
If AI spending slows, Memory makers will eventually need to shift capacity elsewhere. It won't happen overnight, and HBM can't simply become consumer VRAM.
But cheaper GPUs, more VRAM and powerful local AI could finally become a reality. DEMOCRAT-ize access to DGX STATION for the masses 🙏
Maybe we get the good ending after all lol.
Top of the Democratic agenda on day 1 needs to be a pause on advanced AI development, regulation of this out of control industry, and a ban on artificial superintelligence.
1
6
166
Fuck @qrfy_com (qrfy.com) and every other online QR code generator. What a predatory business model.
They turn a basic, free utility into a trap
You are in a hurry, you find whatever website you can find via online search, You give them a URL to generate a QR code. Instead of encoding your URL directly, they route it through their own servers. You download the QR code, print it, and share it, assuming it will keep working.
They deactivate the redirect, making your QR code useless unless you pay.
A QR code is literally just an image containing encoded information. It doesn't need a server, a subscription, or a company to keep it alive.
I used QRFY to generate a QR code linking to important evidence in a Google Drive folder. I sent it to the Police Commissioner as part of an official complaint.
And guess what? The QR code got deactivated.
Imagine sending official documents, printing thousands of flyers, or putting QR codes on product packaging, only to discover that a company can disable them unless you pay a subscription for a fucking url redirect.
The QR code stops working because they've made it unnecessarily dependent on their infrastructure.
This is an artificial problem created to extract recurring payments from users. A free, open standard turned into a subscription hostage situation.
2
3
182
A terrible allocation of scarce resource during a global oil crisis
sam: local inference is a terrible allocation of scarce resources
“a lot of people want local inference, and somehow they assume that the problem is solved now because we have open-source models, we're going to be able to run them on our computers, and so we're not going to need all the stuff in the data centre
but from a global allocation perspective, actually it's quite terrible to have everybody, like, have a bunch of the world's limited supply of DRAM sitting in their own computer and being used five percent of the time
you want maximum utilisation in the cloud if you're trying to deal with this very supply-constrained environment
somehow many people's intuition is the opposite of that”
2
102
Insane to see the impact anthropic has had on local AI even without releasing a single opensource model. Keep it up.
We’ve trained Claude to flag potential misuse in conversations, a technique we call self-assessment. Right after it responds, Claude gives a preliminary safety assessment of the conversation.
Since we began using self-assessment in June, it has surfaced misuse that our other safeguards weren’t designed to catch.
Because Claude has already read the conversation, self-assessment is very efficient. And unlike classifiers, which target a discrete set of harms, it draws on Claude’s own judgment, so it gets more accurate as our models improve.
Read more on our new Safeguards blog:
safeguards.anthropic.com/202…
7
740
Bunty retweeted
OpenAI winning would be bad, Google winning would be worse, and Anthropic winning would be catastrophic. Open-weight models are the only way to ensure that all humanity benefits. If we fail, we will all end up shining the shoes of the San Francisco corpocrats, and that’s the true danger. We really have to get this right.
half my workday is just remembering to do things 💀
reply to that slack thread
check that update
follow up with someone
repeat again till eternity
we built Fluso Agents so you can just tell AI to handle this stuff.
it runs on a schedule, asks when it needs you, and keeps your data private. Powered by open models, hosted in Europe, and built with privacy in mind.
fluso.ai/agents
Fluso Agents are here, and they’re available to everyone!
We started Fluso to make private AI easy to use for individuals and teams, without needing a technical background. Now you can build agents for your work and personal projects by describing the help you need in your own words.
We know setting one up can feel intimidating, so Fluso guides you through the setup. You can keep working with the same agent and adjust its instructions as you go.
Your agents can keep working while you’re away, check back on the goals you’ve set, and let you know when they need your attention. You can also schedule recurring work, so you don’t have to start the same task again every week.
For work involving sensitive data, you choose which apps each agent can use and control what it can read or do. Fluso is built in Europe, powered by open models, and your workspace data is hosted in Europe and kept out of shared-model training.
Immaculate vibes in NK rn
朝鲜最新消息，全球知名媒体BBC、太阳报、联合报、路透社均开始批露朝鲜最高元首、劳动党总书记金正恩已故，临危授命、托孤于其亲妹金于正执掌朝鲜政局辅助其女金主爱，金于正近年来深受金正恩信任，在朝鲜党内、军中已然成为仅次于金正恩的第二号人物，向来以“果断、狠辣、雷霆手段”著称，所以接下来金家王朝是否陷入“权利争夺”未能可知，但外界普遍认为金于正的狠辣手段青出于蓝而胜于蓝，比金正恩更恐怖，朝鲜接下来可能陷入更黑暗的“恐怖时代”。
Readers added context they thought people might want to know
该帖声称BBC、太阳报、联合报、路透社等媒体已披露金正恩已故并详述继任安排，但这些媒体仅报道了源自社交媒体的未经证实谣言。南韩国家情报院评估谣言为假，北韩官媒已发布照片显示金正恩出席近期活动。
bbc.com/news/articles/…
udn.com/news/story/680…
koreatimes.co.kr/foreignaffairs…
en.yna.co.kr/view/AEN202610…
thesun.co.uk/news/40742934/…
1
96
Look at him, that's my VC
Too many people find out on polling day that their name’s missing, or still registered in a city they left years ago.
So I built a free guide: check your name, Form 6 for new voters, Form 8 for movers, and the deadlines that matter.
Open source, no sign-up, 20 languages.
votersrise.pages.dev
1
4
182
Hey @MistralAI @GuillaumeLample @b_roziere @arthurmensch is Large 4 Public Preview covered by commercial TOS §4.3’s Labs/Preview exclusion? With ZDR enabled and training opted out on the console, can prompts and responses still be retained or used for training? Would appreciate some clarification as we are exploring this model.
Support ticket ID: #37633717
4
115
Bunty retweeted
Ten years of Sad Affleck memes and it turns out he'd just been trying to install CUDA drivers the whole time.
Bunty retweeted
The reason why Starshield is so important:
A F-15 vs. a F-117 as seen in the Ka-band at 35 GHz...
Starshield will flood LEO with SAR and infrared reconnaissance sats, everything that flies will get tracked, anything above a certain doppler bin will be processed and classified.
IT GETS BETTER, MUCH, MUCH BETTER.
A 60m generic wide-body, a 37m narrow-body, a Su-34, a F-15, a F-117, a Su-25, a F-16, a Ka-52, and three Shahed drones.
(Look how quiet the F-117 is on the X-band)
@mouthofmorrison @Ascii211 @ElectronicsbyJH
Bunty retweeted
Today, we’re excited to launch Premonition-Exp!
We wanted to bring Jev-like decisions to audio workflows. A request can arrive as a voice note, with the details needed to evaluate it in a document or screenshot. We wanted a model that could work with those inputs together.
So we added audio to DiffusionGemma. Premonition-Exp accepts text, images, video and audio, including mixed inputs in the same request, and returns structured decisions.
For example, you can give it a spoken purchase request alongside an approval policy and ask whether it meets the requirements or needs review.
We trained a 4.3M-parameter projector to connect an audio encoder to DiffusionGemma, keeping both the encoder and backbone frozen.
It’s experimental, and we’re releasing the weights and model code for others to build on.
We just shipped a Jev-compatible API at Prem. Structured decisions from text, images and short video, powered by DiffusionGemma with zero data retention.
Send a customer’s message and screenshot. Ask which team should handle it, whether the customer is blocked and how urgent it is. Get answers your application can act on.
Up to 16 questions in one request. Pick from named options, ask a yes/no question or score against a scale you define. Each answer comes with probabilities.
The interesting part is how it works. We give DiffusionGemma an answer template with spaces to fill. It reads multiple answer positions together in one denoising step, then we turn those scores into typed responses.
I’m particularly interested in document intake and support routing. Check whether evidence is present, flag missing information and send uncertain cases to a person.
Already using the TypeSafe SDK? Change the base URL and model to try it through Prem. It’s live in the playground and API. We wrote up the implementation, tests and limits in the post below.
If you are interested in automating recurring decision in your org's workflow, we would be interested in talking to you.
Jev-compatible decisions are going beyond text.
We’re bringing DiffusionGemma to Prem’s API for questions about text, images and short video, with zero data retention.
See how it works and try your first API call.
2
6
356