Lectura

According to a Semafor report on September 30, the idea for the "White House Accord on Super Intelligence" that President Trump signed with AI executives on September 29 began with a conversation between Meta CEO Mark Zuckerberg and House Speaker Mike Johnson at a state dinner.

The report says Zuckerberg then worked with NVIDIA CEO Jensen Huang on a set of principles, circulated a draft before Tuesday's meeting, and that two other sources confirmed his central role. The accord was signed by Meta, OpenAI, Anthropic, xAI, Google and NVIDIA; at just over 300 words it commits to voluntary internal controls and outside audits.

Meta and the White House did not comment, and the reporting rests entirely on unnamed sources, so the drafting attribution remains unconfirmed by any other channel.

ElevenLabs has completed a $300m employee tender offer at a doubled $22bn valuation.

CEO Mati Staniszewski announced the closing on X on September 30; per Tech.eu, Wellington and T Rowe Price led the offer, with EQT, Goldman Sachs and GIC investing for the first time. In February the company raised $500m in a Series D at an $11bn valuation.

A tender offer is existing holders selling shares to new investors, not new money into the company. Note: Tech.eu's text gives the amount as $300m in one line and £300m in another; the company's own post is the reference.

a16z announced on October 1 that it is leading Armadin's Series B. The round size and valuation were not disclosed.

Armadin was founded by Kevin Mandia, who founded Mandiant, the incident-response firm; his co-founders come from Google's security products and Mandiant's red team. The company's product, Armadin Red, is positioned as an autonomous AI red-team platform that uses agent swarms to continuously test enterprise environments.

The information comes from a16z's own investment announcement. Whether the round has closed, and its size and valuation, remain unknown, and the product's capabilities are the company's description; watch for independent disclosure from Armadin or Mandia.

AI recruiting software startup Metaview announced on September 30 that it closed $60 million in Series C funding, led by Insight Partners with GV, Seedcamp and others participating, bringing total funding to $110 million.

The company says its agents have conducted more than 6 million interviews for over 7,000 companies, and that customers including Deel, Affirm and Replit cut time-to-hire by more than 75%; these figures are company-stated. Its new agent fillmore became generally available the same day, researching candidates, sending outreach emails and scheduling meetings inside Slack.

Metaview plans to grow from roughly 80 staff to more than 250 by the end of next year. Whether agent-run recruiting can genuinely replace human judgment will show up in customer retention and actual hiring quality.

Warehouse robotics software startup Destro AI announced on September 30 that it raised $8 million in a seed round co-led by Base10 Partners and Bonfire Ventures, with CoFound Partners joining. No valuation was disclosed.

Founded less than a year ago, the company's core product MothershipOS is not tied to any single robot maker and assigns tasks across a mixed fleet; a second system, VisionOS, handles perception and grasping on the robots. The founding team came from Hai Robotics, Dexterity and Matic Robots.

The concrete traction is customer Yusen Logistics (Americas): a Pacific Northwest site is moving from a three-robot pilot to a full 26-robot deployment, and a 17-robot pilot is starting in Southern California. Those deployment figures, along with the claim of reaching cash flow positive by year end, come from the company and its investors.

Satlyt, a satellite software company, announced an $8 million seed round led by Houston-based Non Sibi Ventures; the valuation was not disclosed.

The company was founded by Rama Afullo, a Kenyan American engineer who previously worked at Google Cloud and SpaceX's Starlink. Satlyt builds no satellites of its own; it sells software that runs AI models on other companies' spacecraft, describing itself as the "Android" of orbital computing.

Its software has flown twice: earlier this year it ran Google DeepMind's Gemma model on a spacecraft operated by Momentus, and the company says it cut the transmission of onboard software error reports by more than 60% — a company-reported figure. This week its software launches on a spacecraft built by Indian startup TakeMe2Space, with customers including NASA.

Good Tape, a Copenhagen-based AI transcription company, has received investment from Nordic Makers and completed a management buyout that made it an independent company. The funding amount was not disclosed.

According to tech.eu on October 1, the investment was led for Nordic Makers by Alexander Aghassipour, co-founder and former Chief Product Officer of Zendesk, who joined the board. CEO Lasse Finderup bought out the stake held by Zetland, the Danish newsroom it spun out of, saying the spin-out structure had blocked outside investment.

The company also launched Explore, which connects every new recording to everything a user has transcribed before, surfacing contradictions and linking back to the original audio. That description comes from Finderup himself; real-world performance is still unproven.

Good Tape began as an internal tool at Zetland and says it has completed more than 15 million transcriptions in over 100 languages.

BMW announced a restructuring plan on September 30, cutting one fifth of its departments and corresponding management positions by mid-2027, with around 8,000 jobs in Germany affected.

According to Reuters, new CEO Milan Nedeljković said AI will become an important tool for improving organizational efficiency and speeding up decision-making, insisting "this is not a cost-cutting program."

The backdrop: BMW's third profit warning in just over three years on weak China performance. Its core automotive margin stands at 2.3% against a 3%-5% target for 2028, and the stock has lost more than a third of its value over the past year.

According to Sensor Tower, Meta's agentic AI assistant Muse passed 5 million US downloads on September 30, 22 days after launch, topping the App Store free chart for 12 consecutive days.

That is far faster than ChatGPT at 56 days, Grok at 103 days and Claude at 492 days. Sensor Tower counts only 23 US apps ever crossing 5 million downloads within 22 days; before Muse, the only non-game ones were Disney+, HBO Max and Threads.

The growth has a clear paid component: Meta ramped up advertising from September 16, lifting downloads 73% that day, and over the past two weeks Muse took up to 50% of Meta's own ad inventory, shrinking Facebook's and WhatsApp's share. The 5 million figure is a projection by Sensor Tower analyst Kara Lee; the firm's original report is not linked in the source.

Agent long-term memory no longer has to rely on lossy summaries: EnSIMem builds indexes of the form [entity][entity_type][property: value], each preserving source turns and temporal information, with the code open-sourced on GitHub.

Existing memory systems often compress interactions into generic summaries or retrieve anonymous text chunks, making it hard for an agent to identify the right entity, property and evidence. EnSIMem organizes interactions offline into theme-coherent episodes and, online, decomposes each request into evidence requirements and answers from the index.

Xuanyu Meng and five coauthors self-report that the system achieves high answer accuracy while keeping contexts compact; the accompanying repository shows the experiments run on the LoCoMo and LongMemEval long-term conversational memory benchmarks, with ablation scripts included, but the abstract names no baseline or number. The preprint was submitted on September 23, and the performance claim awaits the paper body and independent reproduction.

The value of repairing an agent's failed experience is distinct from the value of reusing it: ThinkingBox's Full mode shows a 44-percentage-point correction gain, yet 12 of its 15-point edge over Skill comes from worse uncorrected performance, not better corrected memory.

Looking only at correction gains can mistake repair value for reuse value, overstating the effect of memory updates.

Experiments by Yanfei Zhang and Xu Lin, submitted to arXiv on September 28, covered 3,300 runs under eleven conditions; the authors conclude memory updates require both a previous-version reference and a fresh-start reference.

All experiments ran on the authors' own ThinkingBox and APEX frameworks, with no third-party replication.

It may be possible to monitor deception and other misalignment in closed models through the internals of an open stand-in, without ever touching the closed model's weights.

A team led by conti published model hermeneutics on LessWrong on September 30: the closed model under study is the author, an open-weight reader reads its outputs, and linear probes are trained inside the reader. In the authors' own tests, a 27B reader probing a 397B author showed an average AUROC gap of only 0.004.

The team reports probes detecting reward hacking, sycophancy and deception in outputs from three frontier models including Claude Opus 4.6, Gemini 3.1 Pro Preview and GPT-5.4, beating direct reader interrogation in roughly half the settings. Distillation lifted the reader's deception-probing AUROC from 0.69 to 0.90.

This is roughly 1.5 months of early work; closed-model internals have no ground truth, so faithfulness is measured only against indirect open-weight baselines, and real audit use awaits independent verification.