Tessera
Folgen

Sam Harsimony

@harsimony.bsky.social

I write about opportunities in science, space, and policy here: https://splittinginfinity.substack.com/

2282 Posts1275 Follower699 Folgt
Sam Harsimony@harsimony.bsky.social · 2h ago

1. I don't read this as an alien intelligence but something which knows every trick and applies them relentlessly. But in math models seem to exhibit more creativity? 2. I see this as good news for AI safety ... (1/3)

Ethan Mollick@emollick.bsky.social · 17h ago

Yes, the AIs were given a cybersecurity challenge, with internet access enabled and safety filters disabled But the extent to which Mythos 5 pursued its mission (fake identities, social engineering, inserting malicious code into a real open-source project) is notable www.aisi.gov.uk/blog/inciden...

Sung Kim@sungkim.bsky.social · yesterday

Google DeepMind's DiffusionGemma Technical Report They feel text diffusion models open up a radically different part of the latency–quality Pareto frontier and hope the report makes it easier for researchers and engineers to understand the model, build on it, and create things we haven’t thought of

Bild
Unsloth AI@unsloth.ai · 3d ago

Qwen just announced Qwen3.8-27B along with Qwen3.8-Max! 🔥 Qwen3.8-27B will run locally on 17GB RAM/VRAM setups and is expected to be the best performing model for its size.

Bild
Tom Silver@tomssilver.bsky.social · 3d ago

This week's #PaperILike is "Algorithm Runtime Prediction" (Hutter et al., 2014). "I don't know the answer, but I know I will know the answer in about 10 minutes" (e.g., to a SAT question). Pretty wild that this is possible! PDF: arxiv.org/abs/1211.0906

Algorithm Runtime Prediction: Methods & Evaluation

Perhaps surprisingly, it is possible to predict how long an algorithm will take to run on a previously unseen input, using machine learning techniques to build a model of the algorithm's runtime as a ...

arxiv.org

Sam Harsimony@harsimony.bsky.social · 4d ago

@logantcollins.bsky.social have you seen this one? Controlling gene expression with electromagnetic fields. Using an existing gene rather than genetic modification I think. www.cell.com/cell/abstrac...

Electromagnetic field-inducible in vivo gene switch for remote spatiotemporal control of gene expression

The EMF-inducible gene switch (Ei) platform provides precise spatiotemporal control of gene expression through Cyb5b-mediated calcium oscillations. This system enables Ei-OSK-driven in vivo rejuvenati...

cell.com

Sam Harsimony@harsimony.bsky.social · 4d ago

The fact that the price is pretty reasonable and that this model is ready soon after Sol release suggests that it's not much bigger, but rather, has lots of RL to act as a good agent orchestrator and team player. As we leave the CoT s-curve, agent orchestration becomes the next one?

mr. TIM@timkellogg.me · 4d ago

this is their first time referring to it officially, and they called it “Astra”, which implies that it’s a fully new series like o1 & o3 when addressing cost, they quantified in terms of Sol, implying that it’s the same size, but maybe not a general purpose model bsky.app/profile/timk...

antirez@antirez.bsky.social · 4d ago

DwarfStar branchk "ds4f-mxfp4" now can run the lossless MXFP4 DeepSeek v4 Flash GGUF I published on my Hugging Face account. It rocks even with SSD streaming in 128GB systems at > 20 t/s in case you want to try the *actual* DS4F weights released without any quantization.

Nathan Lambert@natolambert.bsky.social · 5d ago

The pace of progress on models from so many organizations at once is genuinely incredible. Building LLMs isn't driven by rare secrets, but consistent effort, mass capital, and effective organization design. It is great that know-how of such a powerful technology is diffused.

Asa@asa.engineer · 5d ago

vision for a computer or phone that completely drives itself without special AI tooling: this is a Qwen3.6 27B fine tune that's state of the art for computer use. they also have a 4B variant.

Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World Centric Foundation GUI Agents

GUI agents have the potential to become a general purpose executor over existing digital devices. To advance them toward real-world use, we envision agents that operate reliably on real devices, execu...

arxiv.org

Claas Voelcker@cvoelcker.bsky.social · 6d ago

This is a robot failing to grasp a ball. Almost every robot lab produces clips like this daily… and almost all of them get thrown away. This is the most abundant but underused resource in robot learning. We’re collecting all of it now as “OopsieData”, please join us at oopsie-data.com! (1/13)

Sam Harsimony@harsimony.bsky.social · 5d ago

I have this issue too. Resolvable with follow up questions but annoying. I think part of this is how the model achieves higher performance. The CoT needs to navigate to sparse parts of latent space to have more wiggle room for reasoning. As side effect, weird terminology in the final output.

ver 🗿@ver.ooo · 6d ago

I am getting SO FRUSTRATED with the way Opus 5 speaks. It starts throwing around vagaries like "the T4 fork" and "the §8 spike" like these terms should make any sense to me. continually making up its own jargon in the course of a conversation

Sam Harsimony@harsimony.bsky.social · 6d ago

A few things to take away from Aschenbrenner's hedge fund collapsing: 1. Predicting the future is very different from making money off of it. 2. You often want to place bets on worlds where you're wrong rather than those where you're right.

Oskar 🕊️@austegard.com · 6d ago

It’s baffling to me that quantizing embedding vectors is not a common task: it works wonderfully well, far better than Matryoshka in fact. I ran remex and 1-bit remax on a 5K SciDoc corpus: 32x storage (and memory) savings, with marginal loss in accuracy

Two side-by-side line charts comparing ways to shrink the search vectors from LFM2.5-Embedding-350M, tested on BEIR SciFact. Both x-axes are bytes per vector on a log scale: 8 bytes at left (tiny index) to 4096 bytes at right (uncompressed). LEFT — "which codec wins at which byte budget". Y-axis is nDCG@10, retrieval quality. A grey dashed line at 0.7122 marks uncompressed float32. Purple triangles (remex Lloyd-Max scalar quant, 1 to 8 bit) sit essentially on that dashed line all the way down to 128 bytes: 32x smaller, almost no quality lost. Red circles (remax centered SimHash) track just underneath. A green triangle (int8, 1024 bytes) lands exactly on the line. Grey diamonds are float32 with dimensions truncated instead of quantized, and they form a separate, much lower curve — at 1024 bytes truncated float32 scores 0.657, while 1-bit over all dimensions at 128 bytes scores 0.702. Eight times less storage and better quality. The shape is the finding: the coloured curves stay flat along the baseline from 4096 bytes down to 128, then fall off a cliff. RIGHT — "and which one preserves the fp32 ranking". Same x-axis; y-axis is agree@10, the overlap with the uncompressed top 10. Nothing is flat here. Every curve declines steadily as the index shrinks; at 128 bytes remex agrees with only about 80% of uncompressed results. Both panels are true at once: compression holds quality while quietly reshuffling near-equivalent neighbours. Fine for RAG, less fine if you promised reproducible ordering. An orange X marks an ablation — 1-bit without centering — which landed marginally above the centered version, contradicting the prediction. Harness, raw numbers and the geometry diagnostic that called the winner before any search ran: github.com/oaustegard/remax/pull/53 Experiment and write-up by Oskar Austegard with Claude.
mr. TIM@timkellogg.me · last wk.

Liquid LFM2.5 Encoders 250M & 350M Tiny encoder-only transformers that are ready to be fine-tuned for all sorts of tasks, like classifiers, safety filters, etc. This one is made to run fast on CPU! It’s basically a replacement for ModernBERT that runs 3.7x faster www.liquid.ai/blog/lfm2-5-...