The Stack — September 28, 2026

Fireworks Research ships Ember-1, a reasoning-efficiency play. The model is built on top of Kimi K3 and claims to match K3's output quality while consuming roughly 35–50% fewer reasoning tokens. The underlying problem is real and well understood: reasoning models burn most of their generated tokens on internal "thinking," and in multi-turn agentic loops that prior reasoning gets replayed and re-billed on every turn, so token overhead compounds fast. Fireworks says the work involved 50+ training experiments and 200+ evaluations, validated on public benchmarks and live customer A/B tests, and is serving it as a Research Preview option alongside base K3 with a two-week serverless access window. The efficiency claim is the interesting part; the competitive framing — including a claimed Pareto frontier against GPT-6 Astra and Claude Opus 5 — comes from the vendor's own testing, so treat the head-to-head numbers as marketing until independently reproduced.

DSPy lands in the Elixir/BEAM ecosystem. "Imp" is a full port of DSPy, released as an experimental v0.5 on Hex. It brings signatures, modules, optimizers (GEPA, MIPROv2, SIMBA, BootstrapFewShot), agent loops, retrieval, and MCP/ACP integration to Elixir — with the notable design choice of treating agents as OTP processes, inheriting supervision trees, deadlines, and authorization hooks from the runtime. That's a genuinely different concurrency and fault-tolerance story than the Python ecosystem offers. The author is explicit that the API may change and the optimizers still need large-scale benchmarking, so this is one to watch rather than build on yet.

A deep optimization writeup on llama.cpp's prompt-lookup drafting. The n-gram drafting path got a series of compounding changes: eliminating unnecessary map copies, replacing `std::unordered_map` with `ankerl::unordered_dense` segmented maps, swapping inner hash maps for sorted vectors with branchless binary search, and using a binary-fuse-filter-based immutable map (`constmap`) for the static cache. Net result is roughly 42x faster drafting — up to ~140x with an additional follow-up PR — and up to 2.6x less memory. Acceptance rates are unchanged, since none of this touches the algorithm. If you run speculative decoding in llama.cpp, this is a straight win.

Namespace your Go packages with your own domain. A widely-discussed post makes the case for importing via something like `go.iain.rocks` rather than the git host's path. The argument is coupling: importing through `github.com/...` ties your module identity to a provider you don't control, and migration becomes painful — one cited case had a company paying for three git hosts simultaneously during a move. The post includes nginx and `go-import` meta-tag configuration for setting it up.

A new Markdown editor targeting Mac, iOS, and web. "Beauty" renders Markdown in place (tables, KaTeX, Mermaid), stores notes as plain `.md` files with version history, and offers optional peer-to-peer sharing over WebRTC. The editor itself is free; sync is a subscription. The author notes most of ~1,000 commits went into a contenteditable implementation that prevents the browser from restructuring the document — some of those cross-platform techniques are slated to be open-sourced.

A data-stewardship grievance against NeoVim worth reading on the merits. The account: when NeoVim encountered existing Vim persistent-undo files, it deleted them and replaced them with an incompatible format, breaking undo history in both editors. Maintainers reportedly responded that the persistent-undo format was "unstable" and users shouldn't rely on data preservation in a feature explicitly named "persistent undo." The framing is opinionated and this is a personal account of a past interaction, but the underlying format incompatibility is a known historical issue — and the broader point about features whose names imply guarantees they don't provide is a fair one.

Google is testing in-app purchases inside Gemini and AI Mode in India. Select users can buy a small set of products (smartphones, electronics, accessories) from Walmart-owned Flipkart directly inside the AI interface, with a "Buy" button routing to a Flipkart-branded checkout flow. The test is expected to widen in October ahead of India's festive shopping season. Notably, this appears to use Flipkart's own checkout rather than Google's Universal Commerce Protocol or Google-hosted checkout — a meaningful architectural detail for anyone tracking how agentic commerce actually gets wired up. Google holds a minority stake in Flipkart (~$350M invested in 2024). Rival listings, including Amazon's, show up in the same AI surfaces but without direct purchase options.

Wayve signs Mercedes-Benz. The commercial partnership will integrate Wayve's automated driving tech — Level 2, driver still engaged — into at least one model within two years, following similar deals with Nissan and Stellantis. Wayve is assembling a notable OEM roster without owning a vehicle program.

Fleet and deployment numbers are starting to matter more than announcements. Waymo's ~4,000-robotaxi fleet remains heavily concentrated — roughly 80% in California and Texas — with Texas fleet growth over 49% in three weeks, and the company is expanding to teenage users. Aurora is targeting 30,000+ driverless trucks by 2030 and 200+ by end of year, with CEO Chris Urmson saying the company has "emerged from the building stage." Tesla began delivering its all-electric Semi to customers, though charging infrastructure remains the constraint.

Safety and regulatory items to track. Zoox grounded its Atlanta test fleet after safety drivers were potentially exposed to carbon monoxide, CO2, or hydrogen sulfide — Zoox says only CO2 was detected, and OSHA has opened an inquiry. Comma, George Hotz's startup, faces a federal investigation after five reported crashes involving its aftermarket driver-assistance tech, two of which caused three deaths. VW is reportedly delaying the ID Buzz's US return, while its subsidiary MOIA America partnered with Beep to launch passenger services using self-driving ID Buzz vehicles (Mobileye tech) in Lake Nona, Orlando, with a human operator aboard. Einride will use Nvidia's Hyperion platform for its next-gen self-driving system, and The Boring Company is working on a "simple precursor Hyperloop" between Austin and San Antonio targeting under 30 minutes.

Mobility deals and IPO activity. Carro (SoftBank-backed used car marketplace) is considering a dual listing on Nasdaq and SGX. Spinny (Tiger Global-backed) and PMI Electro Mobility Solutions filed confidentially for IPOs in India. EcoCeres reportedly plans a ~$1 billion Hong Kong IPO. May Mobility announced plans to go public via SPAC merger. Ultraviolette (India electric motorcycles) raised $85 million, adding Intel CEO Lip-Bu Tan as an adviser.

An insurer-funded analysis claims hospital AI use in claims added $942 million in healthcare spending over two years. The Blue Cross Blue Shield Association points to a sharp rise in documented complex conditions without corresponding changes in care delivered. This is a payer-side analysis in an active dispute with providers, so the causal claim deserves appropriate caution — the correlation between documentation patterns and AI adoption is suggestive, not established.

Anthropic's CEO is heading to the White House. Dario Amodei is set to have his first one-on-one dinner with President Trump. The two have been on opposite sides of AI safety debates, and the Pentagon previously designated Anthropic a supply-chain risk over its guardrail attempts — a designation Anthropic is contesting in court. Worth watching whether the meeting signals any thaw in how the government classifies AI labs in procurement contexts.

Meta's consumer AI push continues. At Connect, Meta unveiled the AI agent Muse alongside a Tamagotchi-style device marketed for adults — a deliberate bet on consumer AI while OpenAI and Anthropic concentrate on enterprise and coding. A reviewer found Muse useful for a one-time task (locating unclaimed funds) but characterized it as more of a party trick than a habit-forming tool, and raised the obvious trust question: whether users will hand sensitive financial data to a company whose business model is advertising. That's a reviewer's opinion rather than a verified finding, but the trust concern is structural, not incidental.