Edrak @edrak

Meta’s Omnilingual ASR, just named one of TIME’s Best Inventions of 2026, transcribes more than 1,600 languages. Released 10 November 2025 by Meta FAIR, the system is open source. 500 of those languages had never been transcribed by any AI system before. Meta reports a character error rate under 10% for 78% of the languages it supports. Where at least ten hours of recorded speech exist, 95% of languages clear that bar. With less than 10 hours, 36% still do. The models (300M–7B) were released with recordings in 350 underserved languages, and a new language can be added from just a few examples. Meta FAIR, Omnilingual ASR · released 10 Nov 2025 · TIME Best Inventions of 2026, 15 Oct 2026. Meta-reported results · quality bar: CER under 10 · 22% of supported languages remain above that bar.

Insilico’s AI picked a new target, TNIK, and designed a molecule to block it. In a Phase 2a trial the top dose gained 98.4 mL of lung capacity in 12 weeks; placebo lost 20.3. Phase 3 dosing began on 9 September: 320 patients planned, 47 sites in China, 52 weeks. The drug remains investigational. Even if the trial succeeds, approval is still three to four years away. Insilico Medicine, GENESIS-IPF-3 first patient dosed 9 Sep 2026 · Phase 2a Nature Medicine, June 2025. Rentosertib (ISM001-055), NCT07687459.

AI went back through NASA’s TESS data and found planets that had been missed. Warwick’s RAVEN validated 118 planets (31 new) and flagged 2,000+ strong candidates. Princeton’s search of 83 million fainter stars found 10,091 new planet candidates. Astronomers have confirmed just over 6,000 exoplanets in three decades. Most new finds still need follow-up. The data was already there. AI helped us see it.

This year AI began settling open problems mathematicians had left for decades. In May an OpenAI model disproved Erdős’ 1946 unit-distance conjecture; mathematicians including Noga Alon and Tim Gowers checked the proof and called it a milestone. Google DeepMind’s agent then resolved 9 of 353 open Erdős problems and proved 44 of 492 OEIS conjectures, each at a few hundred dollars, with every proof verified line-by-line in Lean. Tsoukalas et al., arXiv 2605.22763, May 2026.

A man with ALS has used a brain implant at home for nearly two years. It decodes the signals he sends when he tries to speak and speaks them aloud in a copy of his own voice. Over more than 3,800 hours he produced 183,060 sentences (almost two million words) at 56 words per minute. He rated 79% of sentences correct or mostly correct with no researcher present. Card et al., Nature Medicine, 15 Jun 2026. Experimental device, one participant.

On gaps of up to ten characters, Apollo Restore’s top 20 suggestions include the right reading 81% of the time for documentary papyri, 55% for literary and 61% for stone. That is up to 2.6 times the strongest published models, and up to 3.5 times on a length-balanced test. Trained on 600 million words of ancient Greek. Already improved a Herculaneum scroll from 79 CE. McGovern et al., Apollo Restore, arXiv 2609.22455, Sep 2026.

what do we do with the hours the work no longer asks us to spend watching it happen? the rise and fall of the bifurcated screen: on vocation, always-on agents, and the dimensions of screens x.com/i/article/2108…

In a randomised study of 2,383 students, one hour of AI tutoring produced GRE gains statistically equivalent to an expert human tutor: about 14 points, against about 15 with a person and about 8 with no tutor. In 5 of 7 topics, the top AI tutor beat the human average. The cheapest AI tutor to match human results cost 918 times less per point gained. Northcutt et al., StudentBench, arXiv 2609.28470, 23 Sep 2026. Immediate gains, not long-term retention.

In the DRC, the model raised correct cholera hotspot picks by 18% eight weeks ahead. For dengue in Mexico, one-month forecasts improved in up to 72% of active hotspots. In a simulation, it helped reach 5,640 more rural mothers with postpartum depression a year. On heart disease deaths, one month of its data matched census models built on five years. Google Research, 6 Oct 2026. Research results, not deployments.

80% say AI has improved their own productivity. Half say it helps them make better decisions. 44% say it is scaling across the enterprise, up from 38%. 32% expected AI to shrink the workforce. 14% say it did. The top performers redesign the work: nearly three-quarters rebuilt workflows around AI, against about a quarter of everyone else. McKinsey, The state of AI in 2026, 25 Aug 2026. n=1,719.

what are our destinations going to become now that more and more of them are not a lifetime away? the absolute value of now: on ai, mathematics, and distance to the next question. x.com/i/article/2108…

In early 2022, 264 US job titles mentioned AI. By early 2026, 822. About one in 12. New titles include AI Autonomous Truck Test Driver and Physical Therapist (AI Documentation). Searches for AI roles grew elevenfold since ChatGPT launched, and were still under 1% of all searches. Indeed Hiring Lab, July 2026. Standardised titles with at least five postings mentioning AI.

26% of euro area workers used AI on the job in 2024. By 2026, 52%. The typical user saves 3 hours a week. Code saves nearly 8, though only about 8% use it that way. About half of workers, and half of firms, point to training. ECB Blog, 26 Aug 2026. About 20,000 people a month. Self-reported.

Become an AI Company with Edrak.

49% of Copilot chats supported cognitive work: analysing, solving, evaluating, thinking. The rest was people 19%, producing work 17%, finding information 15%. 66% of AI users say it frees time for high-value work. 58% are doing work they could not a year ago. Among the most advanced users, 80%. Microsoft Work Trend Index 2026. Copilot is Microsoft’s own product.

35% of businesses on Ramp paid for AI in March 2025. By August 2026, 56.1%. Venture-backed firms lead at 80%. Private equity, 64%. Everyone else, 45%. Token prices fell 41% from the March peak, to $0.68 per million Ramp AI Index. Card and invoice payments, 70,000+ US businesses. Sample skews tech-forward.

28% of employed Americans used generative AI for work in a given week in 2024. By mid-2026, 39%. Hours assisted rose from 4.1% to 6.3% of all work hours. Hours saved rose from 1.6% to 2.2%. Weekly users saved about 2.2 hours in a 40-hour week. November 2024 survey. St. Louis Fed, FRED Blog, 27 Aug 2026. Bick, Blandin and Deming.

Generative AI reached close to 53% population-level adoption within three years of launch. Faster than the personal computer. Faster than the internet. Estimated value to US consumers rose from $112 billion to $172 billion in a year, up 54%. Median value per user tripled. Stanford HAI, AI Index 2026, ch. 4.