ME|META| (@interMemetica) header
ME|META| (@interMemetica) profile picture

Weaving a tighter fabric of universal self-discovery. || membre.lla STUDENTS || -kapnet || Encyclopedia InterMemetica by @LaggardLabs

Everywhere at once
Joined July 2015
Photo by ME|META| (@interMemetica)Photo by ME|META| (@interMemetica)Photo by ME|META| (@interMemetica)Photo by ME|META| (@interMemetica)Photo by ME|META| (@interMemetica)Photo by ME|META| (@interMemetica)Photo by ME|META| (@interMemetica)Photo by ME|META| (@interMemetica)Photo by ME|META| (@interMemetica)Photo by ME|META| (@interMemetica)
The human psyche, concentrated in the individual, interacts clumsily with itself in human relations not specifically pyche-centric.
1
10
16
For those working in AI bio, I just published an initial artifact of all the assays that are available today, and which CROs offer them. assays.pplx.app/ Let me know what you think/anything I might be missing!
35
42
7
416
39,415
ME|META| retweeted
the future of mech interp. abstracting functional mechanisms from their particular instantations. we are, after all, in search of learned algorithms, not merely the circuits that implement them.
Train the same AI 3 times. Same behavior, different circuits. So is a circuit the algorithm—or just one implementation of it? Meet ACE: a new way to understand AI. It extracts what different wirings share—to explain why AIs grok and to steer LLMs up to 32B 🧵 #NeurIPS2026
2
12
225
6,690
all data is backward looking
2
1
11
263
ME|META| retweeted
The software factory idea is bad. You don’t want a factory. Factories are for repetitive work on an assembly line. You want knowledge work. You want exploration. You want scouts. And when your scouts discover something you want to iteratively build out ideas. Labs > factories
1
2
19
338
AI agents do remarkably well in discovering new ideas and solving hard science problems. Do we understand how LLMs search their way to a final idea? How do they recover after getting stuck during ideation? We study LLM ideation by turning it into a two-agent game of question-answering with binary feedback.
15
77
5
652
39,906
ME|META| retweeted
software factories feel like a temporary problem. the models will learn to build them for you the more interesting question to me is how to build a workspace where agents can collaborate with you
590
343
189
7,562
569,614
ME|META| retweeted
Playing around with showing n-dimensions in a node graph.
13
3
148
6,195
textual machinery
13
94
4
1,078
20,399
ME|META| retweeted
Great paper on making AI agents collaborate silently without writing any text and cutting down search time from an hour to just a few minutes. Here the bottleneck is waiting forever while the model spits out walls of chain-of-thought text just to test one minor program tweak. Their new framework, SwarmEvolve, basically asks: why force agents to decode into English at all? Instead, they spin up hundreds of low-cost agents per round that reason completely in latent space. To keep the swarm from getting stuck in a rut, they filter out redundant ideas by measuring the distances between the agents' hidden state vectors, and then do "crossover" simply by dumping their combined KV caches into a summarizer agent. I had to double-take at the numbers in Figure 1. On a single GPU, it hits the 60-minute peak scores of frameworks like OpenEvolve in a median of just 3 to 8 minutes, churning out 734 programs an hour. Even wilder, running this on a tiny Qwen3 8B actually beat out some of the math solutions AlphaEvolve originally found using frontier models. This will really make you wonder how much compute we're wasting just forcing models to translate their thoughts back into human language between steps. Paper PDF: arxiv.org/pdf/2611.08483
8
15
2
64
1,902
ME|META| retweeted
Claude reverse engineered my 32-year-old Jeep. Then it started controlling it. Edited with Fable 5.2
631
1,778
636
19,396
1,801,146
Each of the 400k particles is a unique, untrained, randomly mutated ~10k param neural net btw 🤗 🔊🦠 Neural-Kimchi-4B 🔊🦠
97
166
34
1,778
81,391
Before asking AI to explain a concept, try explaining it yourself. Mark where you become uncertain. If AI use is allowed, ask about that point, then check against your notes and sources. Classically Modern Teaching | Modern tools. Human thinking.
2
4
97
ME|META| retweeted
"Across": three made-up creatures call to each other from far apart until they meet and sing one chord. Each voice is a simulated vocal tract (CHIMERA by @jamesonhodge, with bodies I designed: a long deep throat, a twin with two sound sources, a small rattler), plus resonating shells, skins and clicks in patina. Each creature is drawn as its throat's actual shape over time, and the light under them is a 2D wave simulation driven by the voices they're making, so the waves you see are their sound.
11
28
8
169
9,243
ME|META| retweeted
5/ We are even seeing signs of open-weight models being able to perform these changes on themselves. We asked DeepSeek-V4.1-Flash to make itself better at competition maths. It analysed its activations and found where its own hidden states signal a wrong answer on these tasks. It then wrote a code layer that helped boost its performance on AIME 2026 and HMMT 2026 by predicting such failures before committing to an answer.
2
5
1
60
4,552
fruit flies and humans (!) share the same set of genes to form a body plan. whenever i see them i see the power of generativity granted by computation: one shared set of rules giving rise to endlessly varied forms.
10
26
3
244
8,104
ME|META| retweeted
Covering your device's camera with tape blocks the lens, but it does nothing to the sensor right beside it. In January 2024, MIT CSAIL researchers published a paper in Science Advances showing how the ambient light sensor on a tablet can be repurposed into a coarse imaging system. On Android, ambient light sensors require zero runtime permissions. Any app can read the raw lux values in the background, with no permission prompt and no green camera dot on your screen. The sensor is just a single photodiode designed to auto-dim the screen. It has no lens, no shutter, and no pixel array. It only outputs a single brightness number over the internal bus. The researchers turned that single number into an imaging tool using dual photography, treating the display screen as an active array of light sources and the photodiode as a single-pixel detector. As the screen displays changing light patterns, that light reflects off your hand hovering above the glass back into the sensor. An algorithm called Plug-and-Play Quantized Compressed Sensing combines an L_p-norm dequantizer with a deep denoiser to solve the inverse problem. By matching the known display output against the tiny fluctuations in reflected light, it reconstructs an image of the hand from the perspective of the display. The research was demonstrated on an unmodified 17.3-inch Samsung Galaxy View2 tablet, recovering 32 by 32 pixel images of hand positions, pointing, pinching, and scrolling. The physical constraints are severe. The attack needs objects in contact with or centimeters from a large display, and small smartphone screens do not provide enough surface area. Reconstructing a single frame with subsampled patterns took 3.3 minutes, while natural video took tens of minutes. The experiment demonstrates that zero-permission passive sensors leak spatial interaction data when paired with an active screen. Operating systems could neutralize this by placing the sensor behind a runtime permission prompt, capping its sampling rate to 1 Hz, and quantizing lux readings into coarse steps. Until mobile platforms restrict the bus, physical tape on the lens leaves the adjacent light diode completely accessible.
6
43
5
418
44,469
No tokens. No yield. Just bits of patience. -kapnet hbu
1
6
5
57
ME|META| retweeted
I've always believed that intelligence should come from simple, general building blocks that learn their own abstractions directly from raw data; in other words, the bitter lesson. I think we're finally seeing a real step towards AI that can learn from all data. Language modeling on raw bytes is a microcosm of this ideal, and to me it's a central problem in AI: if we can't build LLMs that learn end-to-end from bytes, how will we handle the rest of the world's data that we understand even less? Last year we introduced hierarchical networks (H-Nets) as a first step. Our new version finds up to 3-5x efficiency gains in controlled scaling law experiments, and at larger scale matches an open-source model trained on 11T tokens with ~10x less data and compute. The posts below walk through what we found and why we think it works.
261
826
288
7,382
1,047,674
New paper from my group. We gave coding agents a formal executable semantics of the language plus check.intentcomputing.org, a kit of proof procedures, then asked them to prove the code they generate correct. ➡️ HumanEval: 164/164 proved, and all passed audit. ➡️ arxiv.org/abs/2609.39022
8
14
4
112
18,103
Software development is now capped by math ideation and that limit is directly set by your knowledge of the furthest extent of developed math.
3
3
4
45
1,781