Reading

Walapay, a cross-border payments infrastructure company, announced on October 1 that it raised $4.6 million in a seed round led by Generative Ventures.

Walapay uses stablecoins and a single API to handle collections and payouts across currencies and regions, with direct integrations in Latin America, Africa and Asia. The company said the funding will go toward expanding its licensing footprint and deepening banking partnerships.

Valuation and round terms were not disclosed. The company also said its infrastructure could serve autonomous agents making payments on their own, but that remains a plan with no shipped product.

CoreWeave announced its Forge platform at the Fully Connected event, where newly trained model checkpoints can be hot-loaded into live deployments.

The platform includes Agent Lens activity tracing, model distillation, and an RL Rollouts service that updates live models without redeploying the system. Chen Goldberg, executive vice president of product and engineering, said users can start with serverless inference and add pieces gradually.

Customer Cognition's research head Silas Alberti said its Devin agent trains continuously, targeting 99.99% reliability across data centers. TheCUBE, which produced the coverage, is a paid media partner for the CoreWeave-sponsored event, so all of the above is vendor statement.

Volantis announced on October 1 that it has raised an $88 million Series A led by angel investor Lachy Groom and Abstract Ventures, with John Doerr and former Intel AI products head Naveen Rao among the participants.

The startup is building an inference chip with an optical memory interconnect: wiring that reaches beyond 200 millimeters lets it attach more than 220 memory chiplets, which the company says yields over 30 times the memory bandwidth of current accelerators. It plans to ship the A-1, an appliance about a third of a standard rack, next year.

The bandwidth and token-throughput figures are the company's own, and valuation and terms were not disclosed. CEO Tapa Ghosh wrote in a company blog post that the goal is a coding agent finishing a task in 30 seconds instead of 30 minutes — a claim that awaits real hardware.

Microsoft released MAI-Transcribe-2-Streaming, the first streaming transcription model in its MAI family, on October 1. It keeps producing provisional transcripts while the speaker is still talking.

According to Microsoft, the model accepts speech over a WebSocket, returns its first transcript hypotheses within 320 milliseconds on average, supports more than 60 languages with automatic language detection, and is priced at 54 cents per audio hour on the Vercel AI Gateway. The company itself notes that real-world speed also depends on the network connection and the model generating the reply, so the latency figure is not guaranteed in every scenario.

For comparison, the non-streaming MAI-Transcribe-2 released last month costs 10 cents per audio hour and waits until the speaker finishes before processing; the streaming variant costs more than five times as much in exchange for live output. Microsoft also shipped two text-to-speech models the same day, MAI-Voice-2.1 and MAI-Voice-2.1-Flash. Combined with its Mai-Thinking-1 reasoning model, Microsoft now has the components for a full voice agent built in-house, consistent with its stated goal of reducing what it pays OpenAI and Anthropic.

Satlyt, a startup developing AI software for satellites, has closed an $8 million seed round. Per a TechCrunch report on October 1, the round was led by early-stage fund Non Sibi Ventures; valuation and terms were not disclosed.

Satlyt was founded in 2024 by Rama Afullo, previously a product manager on SpaceX's Starlink business. The company lets satellites with GPUs process data onboard, cutting how much must be beamed to Earth. In March, the first satellite carrying its platform reached low-Earth orbit, running Google's open-source Gemma 3 1B model to diagnose software faults in orbit.

The company's own simulated evaluation found the system cut telemetry sent down for analysis by more than 64%; the test simulated a set of common satellite software faults. Another satellite carrying Satlyt's software was scheduled to launch aboard a SpaceX rocket on the day of the report, with NASA using it to test orbital computing methods.

Cloudflare opened its AI Gateway Auto Router to public beta on September 30. Setting the model to cloudflare/auto lets the system pick a model per request automatically.

The router runs a classifier on GPUs across Cloudflare's edge, scoring each request on dimensions including complexity, ambiguity and stakes, then combines those signals with model prices to pick a capable-enough option, accounting for cache-rewrite costs when switching models mid-session.

Cloudflare's own internal benchmark reports: across 291 simulated office tasks, Auto Router scored 86.6% success at $2.10 total cost; Claude Opus 5.5 scored 96.6% at $5.91; GPT-6 Sol scored 84.2% at $2.64. That puts the router at 35% of Opus's cost and 80% of Sol's.

All of these figures come from Cloudflare's own internal benchmark and OpenCode deployment, covering email, calendar and file workflows; whether the savings transfer to other companies' workloads remains to be seen.

The Loughton datacentre the UK government called its "largest sovereign AI datacentre" will miss its planned 2027 launch: the grid operator has told developer Nscale the 90MW it needs will not be available until the early to mid-2030s.

The project is being built in Essex by UK startup Nscale, whose investors include Nvidia and which is preparing for a roughly £26bn listing. When Keir Starmer's government announced the UK AI strategy in 2025, it cited the site as a flagship; in March the Guardian revealed the location was still an operating scaffolding yard.

The delay is not unique. Regulator Ofgem disclosed in June that 315 datacentres were queueing to connect to the National Grid, representing 73GW of demand against a national peak capacity of 45GW.

Limits: UKPN would not comment on the individual project, and the delay timeline comes from reporting; Nscale says it remains committed and is investigating onsite generation and a faster connection.

Tesla's Cybercabs authorized for commercial use in Texas rose from 45 to 169 in their first month, still far below Waymo's 1,154.

According to Texas DMV public records reviewed by CNBC, the wheel-less, pedal-less two-seaters launched in Austin on Sept 3. Paying passengers reported waits of 45-plus minutes and misplaced pickups and drop-offs; wait times have come down in recent weeks.

For comparison, Waymo has 1,154 vehicles authorized in Texas and completes more than 500,000 paid rides weekly across the US. NHTSA opened an audit query into Cybercab compliance after the launch, and Tesla obtained an extension to its response deadline.

Neko Health, a Swedish full-body scan company, has opened in New York, charging $500 for a one-hour scan that is not covered by insurance.

According to a TechCrunch interview published October 3 with early investor Farooq Abbasi of Preface Ventures, the company was co-founded by Spotify's Daniel Ek, has raised nearly $1 billion in total, and has a global waitlist topping 300,000 people. Abbasi said the scan's biomarkers grew from 15 in his first scan to 55 in his latest; these figures come from the company and its investor.

The company still faces scaling and regulatory approval in the US, making the pace of its expansion the thing to watch.

Some image recognition models learn a pile of concepts internally, but no one can say exactly what they recognize. Activation Atlases average the model's internal responses to millions of images and draw them as a concept map, revealing the model's high-level misunderstandings: for example, a baseball photo led the model to recognize a "gray whale" as a "great white shark."

Today, when debugging why an image model misrecognizes something, this kind of visualization is still a standard tool; most later interpretability methods are patches built on top of it. The trigger scenario is a model misjudging something and you wanting to see what it internally took the image to be.

If the model isn't an image classifier, or the images are far from the distribution of common test sets, don't use it to judge model behavior. It draws a low-dimensional projection whose paths don't necessarily correspond to real structure — treat it as a clue, not quantitative proof.

《Activation Atlases》(2019) | Next review 2027-09-20

On September 29, President Trump signed an executive order requiring executive branch agencies to use "Super Intelligence" (SI) in place of "Artificial Intelligence" (AI) in official correspondence, websites, reports and policy documents.

The binding scope is narrow: it covers non-statutory official documents only. Existing regulations, contracts, grants and historical documents need not be altered, and private companies are not bound. The order keeps the existing statutory definition of AI in section 9401(3) of title 15, US Code, and simply renames it.

The real variable is a 60-day deadline: the Assistant to the President for Science and Technology must submit proposed legislative language establishing a federal definition of "Super Intelligence," including an assessment of whether it should supersede the existing statutory definition of AI. If Congress acts, the impact would go beyond wording.

WIRED's commentary reports that no tech executive at the White House lunch objected, but that is the magazine's opinion; the order text itself imposes no requirement on the private sector.

The Financial Crimes Enforcement Network (FinCEN) issued a final rule on August 11 permanently removing the requirement for US companies and US persons to report beneficial ownership information.

FinCEN also announced it will delete ownership records previously filed by US persons, citing the burden on small businesses. Foreign entities remain required to report ownership information for foreign individuals.

Writing on LessWrong, Paul Shearer argues this widens an identity loophole exploitable by AI agents — most US states already register companies without checking who is behind them, while the UK and Australia now verify director identities. That piece is commentary; the rule itself rests on FinCEN's own announcement.