Quick Hits — September 8, 2026
Five items from the evening run: a diffusion model that says it's caught up to the cheap frontier tier, a trade-association split over chip exports, a 2.8-trillion-parameter model running on a laptop, and two smaller notes.
Anthropic is walking out of the tech industry's main Washington trade group over chip export controls. Axios reports Anthropic is ending its membership in the Information Technology Industry Council after the group asked the House and Senate Armed Services committees to strip three export-control measures out of this year's defense bill. Anthropic backs all three, including the AI OVERWATCH Act, which would force the Commerce Department to notify Congress before licensing advanced dual-use chips to adversarial countries. The split is unusual: ITI is where chipmakers, cloud providers, and AI labs normally agree to speak with one lobbyist voice, and Chip makers who sell those accelerators have an obvious interest in fewer restrictions. Take: A lab that argues compute is a strategic national asset cannot sit in a coalition arguing the opposite — expect more labs to hire their own Washington shop.
Inception says Mercury 2.5 is the largest diffusion language model ever trained, and prices it like a cheap model. The company claims a ten-point intelligence jump over Mercury 2 with the same low-latency profile, and lists it at $0.20 per million input tokens and $0.75 per million output, discounted 80% at launch. Customer numbers are the pitch: OpenCall reports median model latency near 170 milliseconds on live phone calls, and Augment Code says moving context compaction to Mercury cut latency 82% — about 150 seconds to 27 — and cost by 90%. Independently published throughput figures run as high as 1,107 tokens a second, and the quality comparison to the cheap frontier tier is the vendor's own claim, not an index. Take: Diffusion models stopped being a research curiosity the moment someone quoted a cost-per-task number.
A 2.8-trillion-parameter model now runs on a MacBook Pro, at roughly one token per second. Deltafin, an open-source project, streams Moonshot's Kimi K3 off four SSDs on Apple Silicon with every one of the model's 16 experts intact — no pruning down to 3 bits, which is how other local builds fit it. The cost is stark: the full install needs about 1.7 terabytes of disk, and the project's own earlier benchmarks on an M1 Max ran closer to several seconds per token, so this is a proof that frontier-class weights are reachable on consumer hardware, not a usable assistant. The interesting part is that the weights stay exactly as Moonshot shipped them, unlike quantized builds nobody has quality-tested. Take: Local inference just crossed "frontier model, unmodified" — just not at a speed anyone would tolerate twice.
Rivian's e-bike spinoff says the AI boom is why your bike is late. Also, the micromobility company Rivian spun out, apologized for months of TM-B delays and blamed "significant supply chain challenges" as a result of the AI boom — specifically, not being able to get enough printed circuit board and integrated circuit material, forcing the team to substitute components and re-test them. The pedal-by-wire DreamRide drivetrain also needs high-precision manufacturing, and the 240W power delivery on the battery exceeds current regulatory guidelines, so new safety protocols have to be written. Take: The memory-and-components squeeze is no longer just a data-center story — it's now showing up in consumer hardware delivery dates.
An open-source library makes non-linear RNNs trainable at transformer-like speeds. ParaRNN, released this week, implements the Newton-iteration-plus-prefix-scan method Apple used to train a 7-billion-parameter RNN, with fused Triton kernels for sLSTM, RWKV-7, liquid CfC cells, and Titans-style associative memory. The author reports roughly 200x wall-clock speedup against sequential training on long sequences and says the iteration budget stays flat out to 131,072 tokens. These models decode one token at a time with constant memory and no growing cache, which is the whole reason anyone still wants them. Take: The efficiency argument for recurrent architectures only works if training isn't the bottleneck — this attacks exactly that.
Sources: Deltafin (GitHub) · ParaRNN (GitHub)
