Technology & Science

SEOUL — Elon Musk said on SpaceX's earnings call Oct. 20 that Terafab, his in-house AI chip project, aims for full-scale mass production in the third or fourth quarter of 2028. Seoul Economic Daily reported that test chips are planned at a research facility in Austin, Texas, around May or June next year. Musk acknowledged the schedule is highly aggressive. Intel is participating and Musk said he would use its 14A process. Talks with TSMC reportedly stalled over who would operate the plant, and Digitimes reported Musk has limited TSMC's role to a possible sublease. Samsung's $16.5 billion contract to make Tesla's AI6 chips through 2033 is not affected immediately, but later orders could shrink if Terafab succeeds. Analysts say early demand could instead lift memory and equipment exports from South Korea.

CARSON CITY, Nev. — Hackers stole about 8.3 million records from a crime-tip platform used by schools and police agencies across the country, Cybernews reported Oct. 25. Nevada is the first state to confirm that its SafeVoice school tip line was compromised. The state blames the vendor, Navigate360, for waiting months to disclose the hack. The report does not detail what the records contain, and the full scope of exposure across other states and agencies has not been established. Tip lines often hold sensitive anonymous reports about students, threats and crimes, which raises the stakes for any exposure of submitter or subject details. Other jurisdictions that use the platform may confirm their own exposure in the coming days.

LiteLLM, maker of a widely used Python package, disclosed Oct. 24 that a master key for its Agent Platform was accidentally posted in a public GitHub repository in late May 2026 and never revoked. According to The Hacker News, an unknown attacker used the key on Oct. 10 to get into the platform and pull out a GitHub command-line OAuth token the service was configured with. That access ultimately let the attacker hijack the release pipeline and push a malicious version of the package. The company has not been reported to have identified the attacker beyond an IP address. Developers who installed LiteLLM during the affected period should check versions and rotate any credentials the package could reach.

The owners of Elvis Presley's music filed suit against Anthropic on Friday, October 23. The plaintiffs are led by an Authentic Brands Group subsidiary, and the case is captioned ABE EPE LLC v. Anthropic. They allege that Claude was trained on 274 Presley compositions without a license or payment. They also allege that it still generates near-copies of the songs. The complaint says that when Claude was asked to recreate songs for a modern audience, it produced versions with the same structure and hooks. Plaintiffs seek statutory damages of up to $150,000 per work, roughly $41.1 million in total. They also want a permanent injunction and destruction of unauthorized copies. An Anthropic spokesperson said the company disagrees with the claims and will defend the case fully. The suit joins claims by Universal, Concord, ABKCO, Sony, Warner Chappell and others. Anthropic has previously argued that training is protected by fair use.

Moonshot AI unveiled Kimi K3.1 Pro on Friday. It is a 3-trillion-parameter, post-trained successor to July's Kimi K3. Third-party tester Artificial Analysis puts it fourth on its Intelligence Index with a score of 55, up from 44. Its Terminal-Bench 4.0 score rose from 13% to 67%, and its cyber-defense benchmark CyberGym-E2E-AA rose to 95%. Moonshot says it will publish the weights under an open license on November 4. If it does, the model would likely become the strongest open-weights system. Until then the model is available only as a service. Results are mixed. On Moonshot's own tables Kimi trails Claude Opus 5.5 and GPT-6 Astra on Terminal-Bench Science. DeepSeek scores higher on one vulnerability benchmark. Artificial Analysis reports a cost per task of about $1.98, roughly unchanged from K3. The headline claims come largely from the lab and one evaluator, so wider replication is still to come.

Data scientist Max Woolf published an analysis on October 26 tracing steep price cuts by OpenAI and Anthropic. He puts reductions at as much as 80% within a few months. He weighs how much reflects genuine inference efficiency and how much reflects competitive pressure, including cheaper open-weight models that limit what leading labs can charge. The same pattern shows up elsewhere this month. Moonshot's Kimi K3.1 Pro is priced aggressively, and DeepSeek's V4.1 Flash is pitched on low serving costs. A caveat in the discussion is that lower per-token prices do not necessarily lower total spending. Agent workloads consume far more tokens, so cost per completed task and overall consumption can move differently from list prices. The piece is one writer's analysis rather than company disclosure, and the labs have not published cost breakdowns.

LONDON — A dermatologist at Guy's Hospital, identified as Lynch, has posted a write-up. It reports that a 10-year-old convolutional neural network, Inception v3, outperformed current general-purpose chatbots at classifying skin lesions. The network was fine-tuned in about 20 minutes on 68,011 images from the ISIC archive. It scored 62% overall, against 42% for OpenAI's GPT-6 Astra and 42 to 43% for Anthropic's Claude Fable 5.1. On melanoma it correctly identified 13 of 30 cases, compared with 8 for Astra and 11 for Fable. The author argues that narrow, purpose-trained image models still beat large language models on specialized clinical pattern recognition. The results are a single author's analysis and have not been peer reviewed. The sample of 30 melanomas is small, and the post is not a clinical trial. It nonetheless adds a data point to the debate over whether general chatbots belong in diagnostic workflows.

BASEL — The Lancet on October 23 published the primary results of REMODEL-1 and REMODEL-2. These are two identical randomised, double-blind Phase 3 trials of Novartis's oral BTK inhibitor remibrutinib against teriflunomide in relapsing multiple sclerosis. In REMODEL-1 the annualised relapse rate was 0.073 on remibrutinib and 0.112 on teriflunomide, a rate ratio of 0.648 (p=0.018). Adverse events leading to discontinuation were less common on remibrutinib (3.1% versus 5.8%). The authors report no liver safety concerns. Bleeding events, mostly mucocutaneous, were more frequent on remibrutinib, while low blood-cell counts were less frequent. The trials enrolled about 2,000 adults. The disability-progression benefit reached only nominal statistical significance in combined analyses. Novartis, the sponsor, has said it plans global regulatory submissions. The comparator was an older oral drug, not a high-efficacy MS therapy.

LONDON — The Lancet published online on October 21 a study of neoadjuvant ipilimumab plus nivolumab in patients with locally advanced colon cancer that is mismatch repair-deficient (dMMR). The journal's summary reports 3-year disease-free survival and 5-year overall survival of 100%. The authors say the findings support the combination as a potential standard of care before surgery for this group. A linked Lancet commentary, titled 'Redefining curative treatment for early-stage mismatch repair-deficient colon cancer', notes that mismatch repair deficiency is found in 10 to 15% of non-metastatic colon cancers. Standard care for resected stage II and III disease has centred on surgery, with consideration of adjuvant therapy afterward. Since the summary gives no details on cohort size or design, the result should be read with caution. A 100% survival figure from a small or uncontrolled cohort would not settle the question without larger comparative data.

DARMSTADT, Germany — ESA's BepiColombo is in the final approach to Mercury. The Mercury Transfer Module separated from the spacecraft stack on Sept. 3. ESA's orbiter has been steering the stack since, and on Sept. 24 it fired its thrusters for 8 minutes 42 seconds, a 11.4 meters-per-second course correction. On Sept. 15 all of the orbiter's instruments were powered up together for the first time since separation. Early analysis shows them in good health, and its science camera SIMBIO-SYS took its first image in space. The spacecraft is due to enter Mercury orbit on Nov. 21. ESA's orbiter and JAXA's Mio orbiter separate on Dec. 9-10, and science operations are set to begin in April 2027. It will be the first mission to orbit Mercury with two spacecraft at once.

Astronomers using the James Webb Space Telescope report what may be one of the most distant supernovas ever seen. The object, designated AT 2026abck, exploded about 1.15 billion years after the Big Bang, when the universe was under a tenth of its present age. Its light left the star more than 12.5 billion years ago. The team believes it caught the blast only days after ignition, which would give an unusually detailed look at the first stage of such an explosion in the early universe. The findings come from a paper by Zhuang and colleagues posted to the arXiv preprint server. They have not been confirmed by independent peer review, and the description of the object as a supernova is still provisional.