Summary
No curated reposts came through this slot, so these 15 posts come from the Following feed, and the best item is a numerics paper. DeepMind's "A Fix for Softmax and Cross-Entropy" finds two rounding bugs in the standard kernels that cause loss spikes and push distilled students away from their teachers, and it fixes them without raising precision. The loudest cluster is harness engineering (cluster of 4): a recycled "End of Harness Engineering" thread, an Anthropic-engineer workshop pitch, a "27B model cracked self-improvement" thread and a Claude Code auto-config trick. The research behind them is real and already covered, but these threads are hype. Two other items are worth a look: EdgeRock-4B, an open model that keeps a fixed-size memory however long its reasoning runs, and a 25-question inference-interview list that doubles as a serving checklist. The rest is courses, tool promos and one non-English joke post.
Posts
- A Fix for Softmax and Cross-Entropy (@askalphaxiv · paper). Barron, Gnanapragasam and Tschannen (DeepMind) name two bugs. "Softmax Collapse" rounds the target logit's gradient to zero when one logit dominates. "Softmax Cancellation" leaves nonzero gradients when student and teacher logits are equal. The fix removes grokking loss spikes (94.3% to 100% in one setting) and stops students drifting from teachers, with no extra precision. It matters for distillation and low-precision training. See knowledge distillation.
- EdgeRock-4B: fixed-size memory for long reasoning (@svpino). An open text-and-image 4B model that keeps a window of recent tokens plus a learned, compressed summary, so memory stays flat as reasoning grows. Predictable memory means more requests per GPU. The costs are lossy recall and the need for a custom serving stack. It's a learned KV-eviction design, see KV cache.
- 25 inference-engineering interview problems (@vishalsingh2972). The list covers sizing a 70B KV cache with GQA, PagedAttention, why FlashAttention is faster at the same math, continuous batching, chunked prefill, prefix caching with cache-aware routing, speculative decoding, FP8/INT4/NVFP4 and H100 vs B200. The engagement framing is bait, but the list itself is a good map of a serving stack. See disaggregated serving.
- Harness hype threads (cluster of 4) (@virgilxbt · @hayatomaruu · @thesupermannx · @Sprytixl). One recycles the "End of Harness Engineering" claim: the loop rewrites the harness against a held-out scoring set, with the score moving from 0.31 to 0.64 over 41 rounds. Another is an Anthropic-engineer workshop pitch on loops, sandboxes and "dreaming" from logs. A third claims a 27B self-improving terminal agent went from 0% to about 20%. The fourth uses Haiku scouts plus Opus to write a repo's
.claude/config. The 10-22 harness-evolution wave already found the held-out exam does most of the work: harness evolution wave. Skip the threads and read agent harness engineering. - "Major contribution for agents" (@JoshARosen). A quote-post endorsing an agent project the capture doesn't show. Click through to read.
- macOS 27 ships a local LLM in Terminal (@dr_cintas). Apple's on-device model can be used via the
fmCLI, with JSON schemas and an OpenAI-compatiblefm serveendpoint. Each session is capped at 4,096 tokens shared between prompt and answer. It's free for summarizing and classifying, too small for code. - Udgam: one person, five open models (@Fintech03). A chat model trained from scratch for $146, a 29.5MB Hindi speech-to-text model said to beat Whisper large-v3 on Hindi, and a 384M reasoning model trained on 17B tokens, all Apache-2.0 with public cost ledgers. The cheap-training receipts are the interesting part. Verify the benchmarks before trusting them.
- LoRA and QLoRA, explained via matrix patterns (@dongxi_nlp). Part three of a Chinese post-training series. It ties rank, low-rank approximation and quantization to one idea: patterns in the weight matrix. A good primer. See quantization.
- Graphify: code knowledge graph for agents (@DivyanshT91162). Uses tree-sitter to build a local graph of calls, imports and inheritance, so agents query it instead of grepping again. It pitches token savings but shows no numbers. Earlier honest evals found these code-context tools worse than their marketing.
- Stanford NLP course "Karpathy would take" (@mentalium_). CS224N lecture links, from word vectors through attention. Skip.
- REA reverse-engineering agents roadmap (@sammy_capital · repo). A promo link dump for an agent toolkit that reverse-engineers apps down to native binaries. Skip.
- Off-topic Russian joke post (@nata3690). Skip.