Counterfactual persona. Genomics for microbes. LLMs for teaching. Occasionally I support my views.

Joined January 2013
Arizona’s Bold Step into AI-Education. So it begins Arizona has approved an unprecedented educational model: a charter school where students will receive two hours of daily academic instruction directed entirely by AI. Unbound Academy, launching next year, is a fully online school for grades four through eight, replacing traditional teaching with AI-driven learning. This model adapts to each student’s pace and ability using tools like IXL and Khan Academy, analyzing emotional cues, time on tasks, and responses to customize lessons. The promise? A learning experience free of frustration or boredom, ensuring optimal engagement and progress. The remaining hours of the school day won’t be idle. Students will engage in life-skills workshops covering financial literacy, entrepreneurship, and critical thinking. These sessions are facilitated by “guides” rather than traditional teachers. This model has roots in Unbound’s private Texas school, where claims of doubled academic outcomes with reduced hours have sparked curiosity and controversy. Inspired by Elon Musk’s iterative approach to innovation, Unbound aims to refine its model for broader adoption across states like Arkansas and Utah. Critics, however, question the reliance on AI, emphasizing the irreplaceable value of human teachers in fostering creativity, empathy, and inspiration. Proponents argue this paradigm shift could address inefficiencies in traditional schooling and better prepare students for a rapidly changing world. Supported by major players like the Bill and Melinda Gates Foundation, Unbound’s AI-driven curriculum might signal the beginning of a significant transformation in education—or a divisive experiment in its accessibility and quality.
62
140
39
883
116,381
I'll get straight to the point. We trained 2 new models. Like BERT, but modern. ModernBERT. Not some hypey GenAI thing, but a proper workhorse model, for retrieval, classification, etc. Real practical stuff. It's much faster, more accurate, longer context, and more useful. 🧵
126
650
116
4,656
437,784
I've created a bug board to fix 🇪🇺 Europe euacc.com/ideas You can post your ideas on how to improve Europe and others can vote on it I've added the first eu/acc objectives that were sourced from X and that everyone voted on here 💪 May the best ideas save Europe!
🇪🇺 I've put the eu/acc goals on euacc.com now All of these are crowdsourced by you We desperately need a ⑂ fork in the road, in Europe too!
219
378
120
3,128
853,366
Introducing a smol Hermes 3 LLM! Hermes 3 3B is now available on @huggingface alongside quantized GGUF versions to make it even smaller. More info and download links here: huggingface.co/NousResearch/… Run Hermes on phones, laptops, and CPUs without sacrificing speed, and may also be a great pairing with 70B for speculative decoding! Hermes 3 3B was built by @teknium, @rogershijin @theemozilla, and @nullvaluetensor Learn more about Hermes, see our technical report, and chat with it now: nousresearch.com/hermes
58
141
30
966
123,921
Schools are going to have to adapt to focus only on things that MUST be learned hands on and in person. Everything else is going to AI.
My children are not homeschooled. But this is 100%. And their teachers told me to stop doing it because my kids “will end up getting bored and not feeling challenged in school.” Currently looking at options.
12
7
3
100
8,968
America needs Education reform. BUT Great Books is for Losers. If you think they (or concepts like "the Canon") provide the solution for restoring American education, you might be a loser too. Some highlights from my latest piece for the American Mind. 🧵
36
26
51
306
153,900
OMG is this a complete AF3-like model that's actually open, not weird pretend open like most of the others?? (I believe Ligo is open but no model weights yet?) "training and inference code, model weights, and datasets" are MIT licensed
Thrilled to announce Boltz-1, the first open-source and commercially available model to achieve AlphaFold3-level accuracy on biomolecular structure prediction! An exciting collaboration with @jeremyWohlwend, @pas_saro and an amazing team at MIT and Genesis Therapeutics. A thread!
4
22
4
143
17,429
Humans are literally walking around with novel, undiscovered proteins inside their gut. There is so much we don't know yet about the living world and that's incredible
A new type of microbe discovered in the human gut! Called obelisks, they can create their own proteins that are entirely new to science. The team named their proteins “oblins.” Scientists scratching their heads about it biorxiv.org/content/10.1101/…
6
46
2
384
35,010
nobody retweeted
Ever think, "This AI companion sounds odd"? You're onto something. Let's explore naturalistic dialogue and how it could change our digital interactions. jxnl.co/writing/2024/08/30/n… I've been focused on dialogue lately. Not the formal kind, but the type you'd hear between friends at a coffee shop. Conversations that flow, full of inside jokes and half-finished sentences that still make sense. Imagine if your AI companion could chat like that.
2
7
65
6,771
Underrated trend
209
625
287
7,366
4,463,447
Meta will *not* release the multimodal versions of its AI products and models in the EU because of an unpredictable regulatory environment. This means that EU users of Ray-Ban Meta won't be able to use the image understanding features. It also means that the EU industry will not have access to future multimodal versions of Llama-3. theverge.com/2024/7/18/24201…
576
802
485
5,735
1,214,962
Occasionally you find science diagrams that would have belonged in House of Leaves -__-
25
657
34
7,009
190,247
Today's the day: Writebook is here! And it's free. once.com/writebook Blogging and posting on social is easy. But why is it so hard to publish a whole book on the web? It’s not anymore. Writebook is remarkably simple software that allows you to publish text and pictures in a simple, browsable online book format. It's perfect for publishing instruction manuals), a collection of short stories, poetry, graphic novels, handbooks, and anything that fits better together than apart. Writebook isn't a service, it's software you download and install on your own server. We've made it incredibly easy to get going — it just takes a few minutes. Even non-technical folks can get it all set up. We'll email you a single command you paste into a terminal and it takes care of the rest. No maintenance required either, it takes care of itself, auto-updates, etc. And it also comes with all the Rails code so you can see how it's built, modify it for your own use, and learn how applications like this are made. So, grab it and dive in! We'd love to see what you publish. once.com/writebook
109
323
113
2,304
406,506
Meet Salesforce Einstein “Tiny Giant.” Our 1B parameter model xLAM-1B is now the best micro model for function calling, outperforming models 7x its size, including GPT-3.5 & Claude. On-device agentic AI is here. Congrats Salesforce Research! Paper: arxiv.org/pdf/2406.18518 Github: apigen-pipeline.github.io
58
200
70
1,334
353,818
Biggest European innovations in the last decade:
287
2,758
283
22,096
1,329,606
🚨 Introducing "ColPali: Efficient Document Retrieval with Vision Language Models" ! We use Vision LLMs + late interaction to improve document retrieval (RAG, search engines, etc.), solely using the image representation of document pages ! arxiv.org/abs/2407.01449 🧵(1/N)
19
79
15
373
78,291
New essay: "Exorcising myself of the Primer" The Diamond Age's "Primer" has long been edtech's most canonical shared vision. I feel the field is haunted by it. It's wrong in many important ways—but we haven't articulated a better one, so we cling to it. I want to transcend it.
12
66
23
530
118,171
Yesterday morning I was happy with myself inferencing llama2.c 10M param model at 18tok/s. This morning people in the PRs are running it at 3000+ tok/s by compiling a little different. Yesterday I kicked off a 44M train run to try slow it down. Now upgrading to GPT-1 sized ~110M.
82
175
35
2,695
481,066
I often get asked about Bayes factors and confidence intervals as alternative to p-values. I don’t think either of these is any solution at all. elevanth.org/blog/2023/07/17…
22
93
20
361
131,262
This is huge
206
2,725
237
28,670
2,309,975