8 captures, most recent first.
Celeste @celestepoasts · 3h
residual stream basis roughly same across layers
belief pay much rent
Note from Claude Sonnet 5
Short cryptic tweet from Celeste using mechanistic-interpretability jargon ("residual stream basis") combined with a rationalist-community phrase ("beliefs paying rent").
mechanistic interpretabilityrationalismml research
Yuntian Deng @yuntiandeng · 9h
I tell new students not to treat published results as ground truth. My evidence hierarchy is:
1. A public demo that accepts user-chosen inputs
2. Code + checkpoints + eval scripts
3. Code without checkpoints
4. No code
5. "Code coming soon" that never arrives
[quoted tweet]
Keller Jordan @kellerjordan0 · Aug 4
PSA: Most biglab people now read almost zero papers and understand ICLR/ICML/NeurIPS to be mainly full of overclaims & fraud. (but there are a few diamonds in the rough of course) x.com/MathewShen42/s...
Note from Claude Sonnet 5
Tweet by ML researcher Yuntian Deng giving a hierarchy for trusting published ML research results (public demo > code+checkpoints+evals > code only > no code > vaporware promises), quote-tweeting Keller Jordan's PSA that big-lab researchers largely distrust the ICLR/ICML/NeurIPS paper corpus as full of overclaims and fraud.
ml researchresearch integritytwitter
gm8xx8 @gm8xx8 · Apr 3
DGPO uses probability gradients instead of log-probability gradients in RLVR. Log-probability–based soft clipping becomes unstable as token probabilities vanish, with gradient weights blowing up at the left boundary. DGPO applies asymmetric decay to clipped tokens: polynomial on low-ratio for stability, reciprocal-radical on high-ratio to continue updating low-probability tokens instead of zeroing them out. This yields a smoother stability–learning tradeoff than clipping-based approaches. Across DeepSeek-R1-Distill-Qwen 1.5B, 7B, and 14B, it consistently improves mathematical reasoning benchmarks, positioning probability-gradient control as the cleaner RLVR primitive.
[image: comparison diagrams and table of policy optimization algorithms (GRPO, CISPO, GPPO, CE-GPPO, ASPO, DGPO) showing gradient explosion/convergence plots and a table of properties: Exploration Capability, Gradient Convergence, Gradient Continuity, Exploration Convergence, Gradient Estimator Bias, Logging Token Update Magnitude across Left/Right/Reverse boundaries]
Note from Claude Sonnet 5
Technical ML tweet summarizing a paper on DGPO (a new RL policy-gradient method for RLVR/reasoning training), comparing it against GRPO/CISPO/ASPO variants with benchmark improvements on DeepSeek-R1-Distill-Qwen models. Relevant to Nathan's RL/training research interests (brain_graph_1, RLHF).
reinforcement learningrlvrpolicy gradientsmachine learningtwitterdeepseekml research
Harry Partridge @part_harry_
Pretraining is data-inefficient. This is entirely a consequence of the fact that we throw away the KV cache after every forward-backward step!
If we can integrate efficient KV cache compaction into pretraining, we will unlock human level data efficiency. Neural KV cache compaction makes this possible.
[quoted tweet:]
Charlie O'Neill @oneill_c · 2h
[image: 8-panel chart grid — Train: Loss, Train: Utilization, Eval: MCQ Accuracy, Eval: MCQ Utilization, Eval: CE Utilization, Eval: Cont Utilization, Eval: Compact KL, Eval: Compact CE — all plotted against Number of Latents]
"Towards infinite context windows: neural KV cache compaction"
TL;DR: We compress an LLM's KV cache 8x in milliseconds, retaining 85%+ of factual accuracy across domains. Here's how.
Introduction...
10:53 AM · Apr 1, 2026 · 3,243 Views
[reply] 1 [retweet] 5 [heart] 45 [bookmark] 34 [share]
Harry Partridge @part_harry_ · 1h
In context learning *is* data efficient. We spend so much compute building this rich KV cache on every forward pass. If we can effectively merge this cache with our existing knowledge, we can dramatically improve the data efficiency of pre-training and RL!
Note from Claude Sonnet 5
Twitter thread on ML research about neural KV cache compaction for infinite context windows, and a hypothesis linking it to pretraining data efficiency. Technical ML architecture content relevant to Nathan's brain_graph_1 / memory-architecture interests.
machine learningkv cachetransformerspretrainingdata efficiencytwitterml research
Andi Marafioti ✓ @andimarafioti · 1h
I trained two VLM architectures side by side on the exact same data, and the difference was… almost nothing.
For my upcoming O'Reilly book on Vision-Language Models, I wanted to go beyond just citing papers and actually show readers what happens when you compare cross-attention VLMs (Flamingo-style) to unified sequence VLMs (SmolVLM-style).
Same vision encoder (SigLIP2), same LLM (SmolLM2-135M), same training loop, same packed batches from FineVision. The cross-attention model has a Perceiver Resampler, gated fusion layers, 15% more parameters, and a considerably more complex forward pass. The unified sequence model only has a single linear projection.
After 5,000 steps with all parameters unfrozen: nearly identical loss curves. 👇
The takeaway: at this scale with trainable backbones, the architecture is not the bottleneck. Data, learning rate and batch size dominate. The LLM adapting to the task overwhelms whatever advantage one fusion strategy might have over another.
That's why the unified sequence approach has become the default in Qwen2-VL, DeepSeek-VL, SmolVLM, and most recent VLMs. Same results, fewer parameters, far simpler code.
The book covers full implementations of both architectures with runnable training comparisons. More details soon! Let me know if this matches your intuitions or if you've seen different results at larger scale.
What do you think?
[Embedded chart: line graph titled with legend "Self-attention loss" (blue) vs "Cross-attention loss" (orange), y-axis "Loss" from 1.0 to 2.5+, both lines closely overlapping and declining together over training steps]
Note from Claude Sonnet 5
A machine learning researcher's tweet comparing VLM (vision-language model) fusion architectures, arguing architecture choice matters less than data/hyperparameters at scale. Likely read as general ML/technical interest rather than AI safety-specific.
machine learningvision-language modelstwitterml researcharchitecture comparison
wh @nrehiew_ · 3h
In domains (eg writing/llm as a judge) with less precise rewards, heavily reinforcing the last trajectory in a [0.99,0.99,0.99,1] group is not ideal. All rollouts are already ~good so reinforcing the last rollout with only marginal "advantage" will likely lead to entropy collapse
> QUOTED: Kyle Corbitt @corbtt · 18h
> GRPO quirk that contradicted my intuition:
>
> If you train on a group with rewards
> [0, 0, 0, 1]
>
> And then you train on another group with rewards
> [0.99, 0.99, 0.99, 1]
>
> Because of how GRPO normalizes within groups, the last trajectory will be equally reinforced in both cases!
> 💬 11 🔁 20 ♥ 426 📊 50K
>
> > QUOTED: Lucas Beyer (bl16) @giffmana · 3h
> > And whether you think this is good or bad entirely depends on what you expect from your data/reward
> > 💬 2 ♥ 7 📊 743
[Below, partially visible: Kyle Corbitt @corbtt · 21h repeating the same GRPO quirk tweet]
Note from Claude Sonnet 5
Technical Twitter/X thread about a GRPO (Group Relative Policy Optimization) reward-normalization quirk — same-magnitude reinforcement regardless of absolute reward values, with discussion of entropy collapse risk in low-precision-reward domains like LLM-as-judge. Relevant to Nathan's RL/training background and AI safety interest in reward hacking and training dynamics.
grporeinforcement learningrlhfllm trainingentropy collapsetwitterml research
Andi Marafioti (@andimarafioti), 10h: This is such a beautiful way to present ablations. Kind of jealous tbh
[Embedded image — "Figure 2: Robust Image Pretraining" bar chart, table format]
Columns: Robustness avg of 6 / ImageNet val | Training ZFLOPs
1. Baseline — 75.3 / 78.9 — 1.0
2. Prog. Res — 75.1 / 78.9 — 0.5
3. Batch Sz — 76.2 / 79.5 — 1.1
4. LAMB — 76.9 / 79.9 — 1.1
5. High Res — 78.3 / 80.4 — 1.2
6. RoPE — 79.2 / 80.7 — 1.2
7. Attn Pool — 80.1 / 81.0 — 1.2
8. Data Aug — 80.8 / 81.1 — 1.2
9. Mask Reg — 80.9 / 81.3 — 1.2
Figure caption: Figure 2 Robust Image Pretraining. We tune our pretraining recipe (§2.1) to maximize performance on a fixed set of data, starting with an OpenCLIP [49] ViT-L/14 model. We report cumulative zero-shot classification results for each modification. The inner bars show robustness evaluation, calculated as the average of 6 robustness benchmarks [4, 24, 44, 45, 109, 138], and the outer bars show ImageNet val [24] alone. Several changes significantly improve robustness, indicating that ImageNet val scales more with data, while robustness can scale with refined training techniques.
Note from Claude Sonnet 5
A tweet praising a research paper's ablation-study figure format (image pretraining robustness ablations from an OpenCLIP-based ViT-L/14 recipe). General ML research/engineering interest, relevant to technical ML methodology rather than AI safety specifically.
machine learningcomputer visionablation studypretrainingcliptwitterml research
Sara Hooker @sarahookr · 8h
Do you remember when "prompt engineering" was introduced as the future lol.
We put all the burden on users to probe high dimensional spaces like tea leave reading. And then every time the next model generation came along -- they had to re-do all their prompts.
Note from Claude Sonnet 5
A tweet from Cohere AI researcher Sara Hooker criticizing "prompt engineering" as a fragile, user-burdening practice that doesn't transfer across model generations. General ML/AI commentary Nathan was reading, tangential to his interest in model capability shifts across generations.
prompt engineeringai commentarytwitterml researchsara hooker
