RuntimeWire

Real-time startup intelligence for the AI economy. Funding, model launches, infra shifts, and founder moves — the signals that matter.

Latest stories — page 20

• MathCode converts plain-language problems into Lean 4 theorems and attempts formal proofs

  Math-AI, the open research community associated with Princeton researcher Yifan Zhang, built a local terminal assistant that stores and reuses successfully proved theorems.

• Stripe agrees to buy OpenRouter for more than $7B

  The deal would pair model routing with the billing stack OpenRouter already uses, months after a $113M round valued it near $1.3B.

• Buf shipped a compiler-backed Protobuf language server for major editors

  Buf integrated its Protobuf language server into the CLI after a limited 2022 proof of concept.

• Portable personal radar claims 6-kilometer drone detection

  The footage shows a rotating flat-panel unit that Steven Cheng says can track targets with extremely small radar signatures.

• Mads Thines launches LoreKit to give coding agents local, inspectable memory

  The open-source tool stores lessons as Markdown, then offers hosted sharing without forcing developers to abandon their local files.

• Pika's April disclosure details one-GPU architecture for lower-cost real-time video

  Founders Demi Guo and Chenlin Meng connect Pika's creator pricing to a one-GPU streaming architecture while keeping model costs private.

• Unslop offers a $19 editor for restoring voice to AI-polished drafts

  Founder Spartak built the English-focused tool after AI grammar fixes made his own posts sound generic; paid rewrites run through Claude.

• Jitpass puts Mac developer secrets behind Touch ID for coding agents

  Jitpass's local-first utility swaps exposed credentials for decoys, then releases real values to approved processes on demand.

• Anthropic nears $7B Decart deal after founders favor it over Nvidia

  Decart's founders and Sequoia reportedly chose Anthropic despite a higher Nvidia offer; most of the consideration would be Anthropic stock.

• Have I Been Flocked lets drivers search 242M Flock audit records

  H.C. van Pelt's public-records tool covers 4.7M plates, while warning that gaps and delays mean a clean result proves little.

• OpenAI lets GPT-5.6 Sol delegate grunt work to cheaper Luna agents

  The new routing can keep Sol in charge while assigning bounded work to faster, lower-cost Luna subagents.

• Duolingo cuts AI video-call cost below 1 cent with open models

  The cost decline is helping Luis von Ahn move conversation practice into Super and reconsider Duolingo's premium Max tier.

• Insilico previews virtual-cell research that models biological age

  Insilico previews a Virtual Aging Cell concept that adds biological age to cell simulations, though the posting date is unclear.

• Waku combines coding-agent CLIs in a native local desktop app

  Version 0.1.0 uses Rust, GPUI and Git checkpoints to combine agent activity without a Waku account or cloud intermediary.

• Startup Spotlight: Sellara is building AI agents for the work after a Wall Street trade

  Sellara's former Citi bankers are building specialized AI agents for trade capture and reconciliation, backed by a16z Speedrun.

• OpenAI benchmarks a 16x load-time gain for giant ChatGPT and Codex threads

  A 741-turn, 231 MB thread loaded in 1.66 seconds in a test, with lower memory growth and 98% fewer requests.

• Printytron turns plain-English requests into dimensioned STL parts without CAD

  Printytron's independent alpha focuses on brackets, adapters and replacement parts, then uses a second AI to inspect each generated model.

• Grok 4.6 tops Newsroom Reliability v0.2 benchmark at 0.78

  In a 50-task run of Newsroom Reliability v0.2, SpaceXAI’s Grok 4.6 ranked first with a score of 0.78. OpenAI’s GPT-5.6 Luna Pro followed at 0.77 with the lowest reported cost among the top three, at $0.0005 per task.

• Netflix tests GenRec on 10% of traffic, reports 0.006% relative lift

  Netflix engineers Ying Li, Arjun Rao and Shradha Sehgal tested an LLM-backed ranker using about 40 times fewer Phase 2 labels than the production model, reporting a 0.006% relative lift in an undisclosed core metric.

• Stephen Cresswell ships Yadda 3 after a one-day Claude-assisted rebuild

  The longtime JavaScript maintainer says Claude handled most of the work, while an existing test suite kept the agent from redefining correctness.

• Sankalp uses Codex loop to place 12th in GPU Mode's B200 QR contest

  Sankalp placed 12th among 183 entrants after using Codex, profiling tools and more than 1,500 submissions to optimize a B200 QR kernel.

• Andon Labs carries out first firing recommended by its AI manager

  Luna needed a human prompt to recover its attendance policy, and people still reviewed and executed the termination.

• Xia Chen ships ThoughtDAG v0.3.13 to expose LLM context

  The open-source desktop app lets users branch, prune, merge and preview the exact conversation history sent to a model.

• Amp holds SOC 2 Type II without requiring pull requests

  Amp, which [LinkedIn lists at 11-50 employees](https://www.linkedin.com/company/amp-code/), uses signed commits, automated checks and agent threads while skipping mandatory pull requests.