./daily — ← back to field notes
The link sheet
One-line pointers to today's local-AI, model, hardware, and healthcare-AI news. No commentary — just the headline, the link, and where it came from.
[ hardware ]9
- Google TPU 8tx Pairs AMD EPYC Turin Hosts with Confidential Computing while TPU 8i Max D… — STH
- b12233 — llama.cpp
- b12230 — llama.cpp
- b12229 — llama.cpp
- b12228 — llama.cpp
- b12224 — llama.cpp
- b12223 — llama.cpp
- Joined the 4x R9700 Club | Initial Benchmarks for QFN 3.8 Below — r/LL
- Dell PowerEdge R770AP Review One Fast Intel Xeon 6 Server — STH
[ new models ]8
- b12236: mtmd: add mtmd_helper_resampler (#31964) — llama.cpp
- b12232: Windows unbuffered model load (#26014) — llama.cpp
- ConceptCountry/qwen-concept-27b · Hugging Face — r/LL
- Nemotron-Personas-Finance-Japan: A Synthetic Dataset Capturing Japan's Demographics and… — HF
- netease-youdao/Confucius4-R2T2 · Hugging Face — r/LL
- b12227: tests : keep dummy ffn.act scales away from zero (#31940) — llama.cpp
- b12226 — llama.cpp
- Your Experts Have Partners. There's HOPE for Keeping Them Together — HF
[ local ai ]12
- Omni Voice is the fastest and best voice clone TTS out there. I've tested them all. — r/LL
- Got another good reminder why local matters today — r/LL
- Qwen 3.8 FN getting stressed from messing up — r/LL
- llm-psicotest: a seeded 38-turn conversation to check whether a quant swap changed behav… — r/LL
- Ming 3.0 TTS sitting alone with no templates — r/LL
- Best Practices for Provisioning Local Inference to a Team? — r/LL
- Muse Ember: Zuck promised to open source muse spark 1.2 but continues to release to more… — r/LL
- From the finest minds of our age to the dumbest monkeys on the planet — r/LL
- Why vector DB metadata filters fail on agent memory write paths (0% raw leak vs 67% leak… — r/LL
- Local first open source alternative to GrokBot — r/LL
- More interesting and unusual Qwen3.5-4B neurons — r/LL
- Any accountants/financial professionals using local LLMs? — r/LL
[ benchmarks & papers ]3
- Top 7 Decision Models Benchmark: Evaluation Across 2,826 Real Tech Decisions — r/LL
- M4 Pro vs M5 Pro (24GB): Qwen 3.8 Flash Next IQ3 with MoEShift at 374 tok/s prefill, 14.… — r/LL
- Multimodal cell foundation models require synergistic integration beyond alignment — NatMI
[ tools ]2
- I fine-tuned a 2B decision model to play Doom on my Mac in 43 minutes (open source, MLX) — r/LL
- Show HN: Uncracked City, detective game where you question LLM suspects that lie — HN
[ healthcare ai ]12
- STAT+: Trump administration removes WuXi AppTec from its Chinese military list — STAT
- STAT+: Bill Ackman’s bold plan to transform brain research — STAT
- STAT+: Hospitals win the midterms, and Chopping Very Strategically — STAT
- STAT+: Heart association meeting’s readout is trending ‘neutral’ — STAT
- STAT+: FDA warns GLP-1 drugs carry potential risk of rare eye condition — STAT
- STAT+: Gene therapy’s long-debated cancer risk materializes — STAT
- STAT+: Eli Lilly’s GLP-1 pill Foundayo helped alleviate sleep apnea in studies — STAT
- STAT+: Pharmalittle: We’re reading about a Novartis heart drug, BIO fighting a Trump pro… — STAT
- STAT+: A new rule to prevent the spread of bioweapons — STAT
- Opinion: STAT+: CVS’ split doesn’t fix its fundamental problem — STAT
- Liver experts finally agree on what acute-on-chronic liver failure is — STAT
- STAT+: Democrats begin asking: How far can we go on health care? — STAT
all days → · 46 links · generated from 9 sources