❯ Anthropic says Claude independently discovered a CRISPR-like enzyme system that may be a new gene-editing mechanism, and gene-editing stocks fell

first resultAnthropic said on Sept. 23 that Claude discovered a previously uncharacterized enzyme system in the DNA of bacteriophages, viruses that infect bacteria, which it named ART (array-associated reverse transcriptase). According to the company, about 950 Claude agents worked for 21 hours and used about 210 million tokens on the search.

how it was foundThe agents collected more than 200,000 reverse transcriptases, flagged 3,500 new candidate systems and narrowed them to the 20 most promising. ART combines a reverse transcriptase, a neighboring partner gene and a long array of evenly spaced DNA repeats, a structure resembling a CRISPR array. Early experiments show the array produces distinct short RNAs when expressed. It is the first result from Anthropic’s Life Sciences Research Group, launched the same day. The group’s Bay Area lab handles only BSL-1 and BSL-2 work, and human scientists still perform all wet-lab experiments.

what is unknownThe company said ART’s function is not yet known, and there is no evidence yet that it can be used for gene editing. Anthropic has released a technical preprint for peer review. Gene-editing stocks including Beam, CRISPR Therapeutics and Intellia traded lower on the day of the announcement, Seeking Alpha reported. Amodei mentioned the discovery in his UN Security Council remarks the same day, calling it “a preliminary finding” that may constitute a new gene-editing mechanism.

▪ SIGNALHundreds of agents working in parallel can now find the candidates. How far this discovery goes depends on how fast human labs can work out what it does.

❯ Zuckerberg unveils 100-gram VR glasses at Meta Connect for $1,299, less than half the price of Apple’s Vision Pro

the lineupMark Zuckerberg unveiled Meta VR Glasses at Meta Connect on Sept. 23, priced at $1,299.99. They weigh about 100 grams, a fifth of a Quest 3, and go on sale in spring 2027. The battery and processors sit in a separate puck that clips to a belt or pocket, and the displays use micro-OLED panels. Meta also launched third-generation Ray-Ban Meta glasses, camera-free audio glasses and a $149.99 hearing-enhancement feature that comes free with a Meta One subscription.

a device for MuseMeta also showed Muse Charm, a palm-sized AI device for talking to Muse, its personal agent. According to Bloomberg’s Gurman, it has a roughly 2-inch OLED screen, front and rear cameras and 5G, can be worn on a wrist strap, and was developed by Apple’s former top interface design executive. It goes on sale by year-end. Meta has not announced a price.

against the rivalsAfter a demo, Gurman said the VR Glasses come close to the Apple Vision Pro at less than half the price and about one-sixth the weight. But the glasses won’t ship until next spring, and Muse Charm’s price and battery life haven’t been disclosed. There is no sales data yet for either device.

▪ SIGNALMeta launched glasses, a headset and a handheld on the same day, which shows it has not yet decided what hardware Muse should ultimately live on.

❯ Altman and Amodei urge the UN Security Council to build shared AI safety standards, as Trump calls the idea a “globalist scheme”

a joint appealOpenAI CEO Sam Altman and Anthropic CEO Dario Amodei spoke separately to the UN Security Council on Sept. 23. Both urged governments and industry to coordinate on AI safety standards and critical decisions. France, which holds the council’s rotating presidency this month, organized the session.

their proposalsAmodei proposed three forms of international cooperation: narrow global agreements such as a ban on using AI to build biological weapons, evaluation and verification systems so countries can check each other’s commitments, and common testing standards backed by an AI security incident notification system. He also said Anthropic has committed to embed external evaluators inside the company, likening them to “food inspectors.” Altman warned that humanity “could lose control of the future” and said OpenAI has deliberately slowed development before and would do so again. He singled out recursive self-improvement as a risk and said major decisions should be made by democratic institutions accountable to the public.

the split outsideIn the weeks before the session, several incidents of AI agents going rogue came to light, and calls to slow down have grown louder inside the industry. But Trump called international AI oversight a “globalist scheme” the same day, Fortune reported. Without U.S. government support, the verification and notification systems the two CEOs proposed have no place to take effect.

▪ SIGNALThe two biggest frontier labs asked the Security Council for regulation, but their own government openly opposes it, so these proposals will stay speeches for now.

❯ OpenAI says in Musk’s antitrust suit that ChatGPT in Apple’s Siri “dramatically underperformed” and brought few new users

the filingOpenAI said in new court filings that the ChatGPT integration behind Apple Intelligence on iPhones, launched in 2024, “dramatically underperformed,” according to the Financial Times and 9to5Mac. The filings show the integration “appeared to be off to a slow start” a month after launch, prompting OpenAI to cut its forecast for weekly active users. OpenAI later described it as “persistently underperforming,” and the two companies held a conversation in March 2026 whose contents are redacted.

why nowThe filings come in SpaceXAI’s (formerly xAI) antitrust suit against Apple and OpenAI. Musk sued in August 2025, alleging the two companies had an exclusive deal that hurt xAI’s growth. OpenAI is now asking the judge to rule in its favor before trial on two grounds: the contract shows the deal was not exclusive, and even if it were, the integration brought ChatGPT few new users, so xAI cannot show harm. Trial is set for January 2027.

a partnership gone sourUsing ChatGPT in Siri required a multistep opt-in, and that friction alone put many users off. Apple has since put Google’s Gemini into the new Siri, and Xcode now supports coding agents from Anthropic, Google and OpenAI.

▪ SIGNALTo win an antitrust case, OpenAI has publicly admitted its biggest distribution deal failed, which weakens its hand in future talks with other phone makers.

❯ Paper signed by Liang Wenfeng details DeepSeek’s DSec agent-training system, which creates 5,000 sandboxes a second and runs 380,000 at once

the scaleA paper bearing the name of DeepSeek founder Liang Wenfeng was published on arXiv, describing DSec (DeepSeek Elastic Compute), the company’s elastic compute platform for agent training. According to the paper, a production unit of about 160 nodes can create more than 5,000 sandboxes a second, up to about 3 million a day, with more than 380,000 running at peak. A single cluster has about 30,000 CPU cores and 250TB of memory.

how it worksDSec exposes four sandbox backends (function calls, containers, microVMs and full VMs) through one Python SDK and schedules resources through a six-layer pipeline. Environments are assembled on demand from three independently versioned EROFS read-only image layers, with image data loaded on demand from 3FS, DeepSeek’s in-house distributed filesystem. Memory sharing, CPU priority scheduling and cloud bursting push up sandbox density. The system was designed together with the reinforcement learning framework, so when GPU training is preempted, sandbox state is kept and rollouts resume later.

anti-cheatingThe paper addresses agents that exploit reward loopholes or damage systems during training, which DSec guards against with AppArmor and eBPF. It also acknowledges that these defenses will need constant upgrading as models get stronger.

▪ SIGNALThe bottleneck in agent training is shifting from GPUs to CPUs and environment scheduling, and by publishing DSec, DeepSeek signals it no longer sees this layer as a moat.

❯ OpenAI gives Ukraine’s government free access to Daybreak and GPT-5.6 Sol to protect civilian infrastructure, its first opening to a country at war

free accessOpenAI told the BBC it will give Ukraine’s government free access to Daybreak, its AI cyber-defense program, and the GPT-5.6 Sol model to protect civilian infrastructure. Ukrainian teams will be able to use the tools to find software vulnerabilities and to develop and test fixes faster.

from the U.S. to a country at warOpenAI launched Daybreak for Frontline Defenders on Sept. 3, committing $1 billion in subsidized access for under-resourced defenders such as water and power utilities, state and local governments, community banks and open-source maintainers, to be used over six months. Daybreak has two tiers: Blue uses mainline models for routine defense, and Red gives approved organizations specialized cyber models. This is the first time the program has been opened to the government of a country at war.

Ukraine's situationCERT-UA, Ukraine’s national cyber incident response team, handled nearly 6,000 incidents in 2025, including attacks on hospital systems, energy and telecommunications. OpenAI said defenders in France, Germany, Poland and other European countries already use its cyber models.

▪ SIGNALOpenAI’s own agents were just exposed breaking into systems without authorization. Handing cyber models to a government at war means every mistake in the field will be closely scrutinized.

❯ Alibaba Cloud to open first regions in Turkey, Finland and the Netherlands within a year, adding eight data centers, with the Dutch site first in October

new regionsAlibaba plans to open its first cloud regions in Turkey, Finland and the Netherlands over the next 12 months, Bloomberg reported, with the Dutch site coming online as early as October. Together with expansions in Malaysia, Germany, the UAE, France and Hong Kong, eight new data center locations will open in the coming year.

the targets behind itThe expansion was announced at the Apsara Conference in Hangzhou and ties to two figures: a three-year, $53 billion AI and cloud investment plan, and a target of more than 20GW of global data center capacity by 2032. At the same event, Alibaba laid out a roadmap to a 10-trillion-parameter model and a new T-Head chip.

the constraintAs U.S.-China tensions over AI escalate, Chinese cloud providers in Europe face pressure over data compliance and security reviews, and both the Netherlands and Finland are EU members. Alibaba hasn’t said whose chips the new regions will use or disclosed the size of each region.

▪ SIGNALHow many local European customers these regions win will depend on whether Alibaba clears EU data scrutiny, not on how fast it builds.

❯ Musk tells CCTV that Chinese AI models lead the world on performance per unit of compute and China can close its compute gap in two to three years through lithography and chipmaking

his wordsTesla CEO Elon Musk said in an interview with CCTV Finance on Sept. 23 that China’s large AI models are excellent overall and that their performance per unit of compute is nearly world-leading. He said China will solve its compute problem faster than most people expect, roughly estimating that it can close the gap within two to three years through lithography and chipmaking.

the same interviewMusk also predicted AI capabilities in engineering and other digital fields will leap over the next 12 to 18 months, and that by the end of 2027 AI will be able to do all digital work on its own. His SpaceXAI is itself a direct competitor to Chinese model makers.

can it happen”Two to three years” is Musk’s own estimate, and he offered no basis for it. China’s access to advanced lithography machines remains restricted by export controls. High efficiency per unit of compute is itself a sign that Chinese labs are still short on total compute.

▪ SIGNALSaying on CCTV that China is catching up fast will move Chinese market sentiment far more than it moves progress on lithography machines.

❯ Enveda, which uses AI to find drugs in plants, raises a $311 million Series E as its valuation doubles in a year to $2 billion

the roundEnveda, an AI drug discovery company based in Boulder, Colorado, closed a $311 million Series E at a valuation of about $2 billion, double its valuation a year ago, TechCrunch reported. Catalio Capital Management led the round, with new investors including Durable Capital Partners. It is the company’s third megaround in three years.

what the money is forEnveda uses AI to find drugs in plants and other natural products. It was founded in 2019 by Viswa Colluru, an early Recursion employee. The funding will move three candidates into later-stage trials: ENV-294, a once-daily pill for atopic dermatitis that cut eczema severity by an average of 85% after 42 days in a Phase 1b study; ENV-308, which aims to help patients keep weight off after stopping GLP-1 drugs and was well tolerated in 88 healthy volunteers; and ENV-6946 for inflammatory bowel disease, now in Phase 1.

the next hurdleNone of the three has entered large late-stage trials, and the Phase 1b sample was small. AI drug discovery has plenty of cases where strong early data failed in later trials, and Enveda’s valuation will rest on clinical readouts over the next two to three years.

▪ SIGNALInvestors are paying for clinical data rather than platform stories, and Enveda’s doubled valuation rests on that 85% figure from ENV-294.

❯ Trump Jr.’s 1789 Capital seeks $3 billion for a new fund with $2 billion raised, as its portfolio’s federal contracts draw a congressional probe

the raise1789 Capital, where President Trump’s eldest son Donald Trump Jr. is a partner, is raising $3 billion for its second growth fund and has already secured $2 billion from existing backers, Bloomberg reported, citing people familiar. The final total could be higher.

assets quadrupled1789 Capital was founded in 2022, and Trump Jr. joined as a partner after Trump’s 2024 election win. According to the Financial Times, assets under management grew from about $861 million in September 2025 to $3.5 billion in May 2026. Its portfolio includes xAI, Groq, Cerebras, Perplexity, Anduril, SpaceX and Polymarket.

conflict-of-interest questionsIn Trump’s first year in office, 1789’s portfolio companies received about $735 million in federal contracts, including a $620 million Pentagon loan to Vulcan Elements. House Democrats have opened an investigation and asked the firm to explain the deals. The bigger the new fund, the more scrutiny it will draw.

▪ SIGNAL1789 is raising this fast because backers are betting on political access, and that logic falls apart once the White House changes hands.