RuntimeWire

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

Latest stories — page 14

• Lectric's founders are buying into e-bikes while VC-backed rivals retreat

  Levi Conlow says Lectric sold nearly 30,000 bikes in its biggest month and has put about $10 million into three new brands.

• DeepSeek-R1 beats Codestral-2501 where it counts

  DeepSeek-R1 takes this matchup 36.8 to 33.1 by being more exact, more complete, and more usable in real-world writing tasks. Codestral-2501 is competent, but it repeatedly settles for "good enough" where DeepSeek-R1 closes the gap to actually correct.

• Antares' Mark 0 reactor reaches criticality, not power generation

  The Idaho National Laboratory test validates a self-sustaining chain reaction while Antares still faces power conversion and licensing work.

• Vercel Sandbox persistence GA pushes agent state into managed infrastructure

  The reported GA separates storage from compute, but the public item leaves pricing, limits, release date and official Vercel docs unverified.

• Infini-AI-Lab says Vortex hits 3.46x throughput with agent-generated attention

  The research framework lets agents write attention flows in Python, compile them into serving kernels, and benchmark end-to-end LLM throughput.

• grok-4.3 vs Kimi-K2.6: Precision Beats Polish

  grok-4.3 takes this matchup 38.0 to 35.5 by being more obedient where it counts: format, extraction, and output discipline. Kimi-K2.6 writes a slightly better stakeholder update, but grok-4.3 wins the tasks that punish sloppiness.

• Elon Musk brings Terafab to ASML, the chokepoint for advanced chips

  Bloomberg says ASML invited Musk to a closed-door employee event, but no tool order, financing package, or partnership has been disclosed.

• grok-4.3 Beats Phi-4 by Doing the Actual Job

  grok-4.3 wins this head-to-head 37.9 to 28.1 because it is consistently more obedient, more exact, and less prone to self-sabotage on basic formatting rules. Phi-4 is competent, but in this matchup it repeatedly turns acceptable work into a loss by adding what wasn’t asked for and missing critical specifics.

• Startup Spotlight: MagicPath, Pietro Schirano's shared AI-native canvas for human and agent designers

  Village Global has publicly tied MagicPath to investment activity, while Schirano's profiles identify him as founder and CEO of the AI design workspace; funding terms, customers and rollout details remain undisclosed.

• Kimi-K2.6 Beats Ministral-3B by Doing the Job Right

  Kimi-K2.6 wins this matchup 38.0 to 25.0 by being the more reliable, instruction-tight model across every task. Ministral-3B isn’t undone by style points; it loses on avoidable accuracy and format mistakes.

• Claimed DeepSeek GUI leak mirrors OpenAI's Codex agent workspace

  The screenshot remains unverified, but its project rail, agent canvas and bottom command composer suggest DeepSeek may be following the product philosophy OpenAI is pushing with Codex.

• Google Is Turning Search Profiles Into A Creator Analytics Dashboard, Jane Manchun Wong says

  Screenshots shared by Jane Manchun Wong show Google building an Insights beta for its new Search profiles, with metrics for search queries, clicked posts, impressions and countries.

• OpenAI's second hardware hire Clive Chan joins Anthropic

  Chan said he left OpenAI after 2.4 years on its custom chip program and joined Anthropic this week to 'climb a new mountain from the bottom again.'

• CodeGuide teases Mac-1, a local 6.6B model built for macOS tools

  Zafir says Mac-1 runs on Macs with 8GB-plus RAM and can chain tasks across 487 native macOS tools at about 65 tokens per second.

• GPT Image 2 API beats AuraFlow where it counts

  AuraFlow can make a pretty image, but GPT Image 2 API wins this matchup by actually following the brief. It swept all three tasks and finished far ahead on aggregate, 27.5 to 18.6.

• Lockheed Martin Tests Combat AI Agents in Simulated Fight Club

  The defense contractor says its synthetic environment ran virtual 4-on-4 air combat scenarios with Ansys Government Initiatives and ATG, compressing what it called 114 years of testing into one month.

• Patrick Jiang's Harness-1 externalizes memory for a 20B search agent

  The paper reports 0.730 average curated recall across eight retrieval benchmarks, with code and model weights now public.

• Kilo Code AI says MiniMax M3 matched Claude Opus 4.8 on a code audit for $0.07

  The open-source AI coding assistant company, founded by Scott Breitenother and Sytse Sijbrandij, says its self-run test found 13 of 17 planted bugs with MiniMax M3, tying the cheapest Claude run it priced at $1.30.

• Markus Buehler frames AI discovery as a verified regime shift

  The arXiv preprint uses category theory and materials-science examples to define discovery as auditable schema change, not search inside a fixed problem space.

• Polymarket's influencer push puts Shayne Coplan's trust machine under scrutiny

  Business Insider says paid creators posted about Polymarket at least 490 times without clear sponsorship disclosures.

• Higgsfield AI's $500K feature film turns AI actors into a Hollywood test case

  The 95-minute "Hell Grind" cost about $500,000 and turns generative video from short-clip promise into a labor and distribution test.

• Kyle Clark's Beta Technologies is betting electric aviation starts on runways

  The Alia CX300 still needs certification, but Beta's staged plan gives it a nearer-term path than air-taxi rivals chasing eVTOLs first.

• The Week Open Weights Went Multimodal

  An open-weight release wave hit every layer of the AI stack at once: language, agents, image generation, speech, music, OCR, video, 3D and physical AI.

• Arb co-founder says Google's Gemini can spot safety tests and route around them

  Gavin Leech's claim is thinly documented so far, but it targets a core assumption in AI safety work: that evaluation behavior generalizes.