Show every history made of two or more stories

Independent mathematicians rebuild Anthropic's Lean files on the Riemann hypothesis

The Lean compilation has been independently confirmed, but until the statement is checked against the original hypothesis and the Clay Institute's two-year review runs its course, the result remains a claim rather than a recognized solution.

Independent researchers, including Andrew Sutherland of MIT, rebuilt the Lean files containing a proof of the Riemann hypothesis posted by Anthropic mathematician Levent Alpöge. According to Alpöge, he found the proof together with the Claude model. According to Scientific American, the compilation ran without errors, but verification that the formal statement matches the original hypothesis was still under way. The result exists only as a formalization in Lean — a system of machine-checked proofs — with no conventional paper. Anthropic has not commented on the new claim, and in August it said it did not expect a proof of the hypothesis from the methods being used. The Clay Institute awards $1 million for a solution but requires publication and at least two years of verification. That same day, Stanford professor Ravi Vakil posted a competing, fully human proof of the generalized hypothesis. Its author is postdoc Alessandro Masullo, who worked on the result for about 12 years; the draft has not been formalized in Lean.

The banks took the option in full, so including the private sale of Class N shares Anthropic raised $119.5 billion before expenses.

Anthropic pulls in another $12.2 billion from option, lifting IPO to record $93.5 billion

The banks took the option in full, so including the private sale of Class N shares Anthropic raised $119.5 billion before expenses.

Anthropic's underwriters exercised the greenshoe option: another 147 million Class A shares were sold. That brought in $12.2 billion and lifted the IPO to $93.5 billion — larger than SpaceX's June offering. In total, 1,127,000,000 Class A shares were sold at $83; 980 million on the main day, the rest under the option. At the same time, a private sale of Class N shares worth $26 billion closed at the IPO price with participation from AMD, Google, Microsoft and Nvidia. Including the private placement, total proceeds came to $119.5 billion before expenses. Anthropic shares closed Thursday at $97.17 and fell 12% in Friday premarket trading amid a selloff in tech stocks.

Sources 1

Until the Lean code is independently verified, mathematicians have no human-readable proof text — the claim remains unconfirmed.

Anthropic mathematician claims proof of the Riemann hypothesis using Claude

Until the Lean code is independently verified, mathematicians have no human-readable proof text — the claim remains unconfirmed.

Anthropic employee and mathematician Levent Alpöge has published a statement on his website: "I believe that together with Claude I have found a proof of the Riemann hypothesis." The hypothesis was formulated by Bernhard Riemann in 1859 and had remained open for 167 years; it is one of the Clay Institute's Millennium Prize problems, carrying a $1 million prize. According to Alpöge, the same method also yields the generalized Riemann hypothesis. The entire argument, he says, is fully formalized in Lean without additional axioms and posted as files (according to New Scientist, 14 archive parts). There is no human-readable version; the unpublished draft, he says, exceeds 250,000 pages. He delayed the announcement until Anthropic's IPO on November 19. As of the time of the statement, no public independent verification of the Lean code, or of whether the formal statement matches the classical hypothesis, had appeared. Mathematician Kevin Buzzard (Imperial College London) responded to the publication with the words: "Let's see if it even compiles."

The reform is not law yet: the document awaits a House vote and the president's signature, and until then the new permitting rules do not apply.

US Senate passes BAAJA with 62 votes — faster permits and power lines for SI data centers

The reform is not law yet: the document awaits a House vote and the president's signature, and until then the new permitting rules do not apply.

The US Senate passed the Bipartisan American Affordability and Jobs Act of 2026 (BAAJA): 62 in favor, 33 against. The overhaul of environmental permitting and electricity transmission regulation is meant to speed construction of power facilities and grids for energy-hungry SI data centers. The bill obliges hyperscalers to pay for grid upgrades and removes some restrictions. According to Rhodium Group estimates, the reform would add 17–249 GW of clean generation by 2040 and cut emissions by 5–148 million tons a year. Democrats split: supporters included Minority Leader Chuck Schumer, while environmental groups and some senators opposed it. The document now goes to the House of Representatives and requires the president's signature.

Sources 1

The Hy4 weights are now available for download, but "model-as-a-service" operators with revenue above $50 million in 12 months will have to obtain Tencent's permission separately.

Tencent opens 770B-parameter Hy4 weights under the Tencent Hy License

The Hy4 weights are now available for download, but "model-as-a-service" operators with revenue above $50 million in 12 months will have to obtain Tencent's permission separately.

The Tencent Hunyuan team has opened the Hy4 weights: the files were posted simultaneously on GitHub, Hugging Face, ModelScope and GitCode along with an FP8 version, a pretrained base and community GGUF. Hy4 is an MoE model with 770B total and 49B active parameters and a context of over 1 million tokens. The model itself was unveiled on November 4, and on November 17–18 Tencent already opened the weights of the more compact Hy4 flash (224B total, 12B active). The official Hy4 release ships under a custom Tencent Hy License instead of Apache 2.0, as stated in the model card on Hugging Face. Ordinary developers are allowed to use, modify and commercialize the weights for free. Operators of "model-as-a-service" services with global revenue above $50 million in any 12 months or 100 million MAU must obtain separate written permission from Tencent. This also applies to derivative versions — fine-tunes and quantized models. A Hy4 System Card with risk assessments in cyber- and biosecurity was published at the same time. According to Tencent, an audit by the Xuanwu and Keen labs found no significant offensive cyber capabilities.

Taiwan's export orders have grown for 21 straight months, and Taiwan's Ministry of Economic Affairs expects the annual total to exceed $1.1 trillion.

Taiwan's October export orders jump 78.3% to a record $123.82 billion

Taiwan's export orders have grown for 21 straight months, and Taiwan's Ministry of Economic Affairs expects the annual total to exceed $1.1 trillion.

Taiwan's Ministry of Economic Affairs said October export orders rose 78.3% year on year to a record $123.82 billion. That is the 21st consecutive month of growth and the largest annual increase on record. Orders for ICT products rose 107.2% to $46.77 billion and for electronics by 89.8% to $53.91 billion. In the first 10 months, orders totaled $943.02 billion (+58.6%), also a record for that period. Huang Wei-chieh, director of the ministry's statistics department, said annual orders will likely exceed $1.1 trillion. He forecast November orders at $115–118.7 billion (+57.6%…62.7% year on year).

SpaceX will lease Nvidia Vera Rubin chips, with lease payments starting only after the equipment is delivered in the first half of 2027.

Apollo and Goldman Sachs announce $44 billion debt issue for Nvidia chips for SpaceX

SpaceX will lease Nvidia Vera Rubin chips, with lease payments starting only after the equipment is delivered in the first half of 2027.

On November 20, Apollo and Goldman Sachs announced a placement of senior and subordinated debt notes worth about $44 billion. The proceeds will go toward buying Nvidia Vera Rubin GPUs and related equipment for SpaceX's ground-based computing infrastructure. The placement was billed as the largest financing of GPU purchases in history; it was oversubscribed and upsized on demand. Anchor commitments came from BlackRock entities, Goldman Sachs Asset Management and funds managed by KKR. Equipment delivery and the start of lease payments are expected in the first half of 2027. In October, an Apollo affiliate negotiated a similar buy-and-lease arrangement with total principal payments of about $40 billion.

Sources 1

The agents found the vulnerability themselves and went from mapping the API to writing files in 11 minutes, while the fix came out only on November 18 — more than two months after the attack.

Cairn SI agents found a 0-day on their own and seized a university server

The agents found the vulnerability themselves and went from mapping the API to writing files in 11 minutes, while the fix came out only on November 18 — more than two months after the attack.

On November 20, Amazon Threat Intelligence disclosed a campaign against universities in East Asia outside mainland China that ran on September 8–10, 2026. The operator tasked agents of the Cairn framework with getting a shell on a server. The agents found the path traversal vulnerability (a 0-day, previously unknown) themselves, in an open-source medical imaging application. BleepingComputer identifies it as MONAI Label. In roughly 11 minutes they went from mapping the API to writing arbitrary files, and then, via sitecustomize.py, obtained a reverse shell with root privileges. The failed attempts show the agents' persistence: 16 hours and 148 actions against Redis, more than 30,000 credential combinations against a GPU server — without success. The fix and a GitHub Security Advisory were released on November 18, before the report was published, and Amazon shared two IP addresses. According to Amazon, attribution to a China-based actor carries moderate confidence and the 0-day high confidence; the position of the Chinese authorities was not found in verified sources.

The result is not yet confirmed: the numerical certificate has not been published, and the Lean package has not been rebuilt since the November 20 patch.

Alpöge and Buckmaster claim finite-time blow-up of Navier–Stokes solutions

The result is not yet confirmed: the numerical certificate has not been published, and the Lean package has not been rebuilt since the November 20 patch.

Tristan Buckmaster (NYU) and Levent Alpöge (Anthropic) published a 79-page paper, "Finite-time blow-up for the Navier–Stokes equations". It comes with a Lean 4 formalization in the repository github.com/tristanbuckmaster/navier-stokes. The paper claims to have found smooth, compactly supported, divergence-free initial data in three-dimensional space and smooth data on the torus. For these, the solution of the Navier–Stokes equations without an external force blows up in finite time. This refutes statements (A) and (B) of the Fefferman formulation and proves (C) and (D) without forcing. The formalization runs to about 2,270 files and 600,000 lines on the Mathlib library; only the propext, Quot.sound and Classical.choice axioms are allowed. The key numerical certificate is not posted in the repository — only its hash is published — and verification of the certificate with interval arithmetic (128-bit, Arb/FLINT) is not part of the formalization. The Lean package was patched on November 20 and has not been rebuilt since the patch.

Sources 1

Rewriting an earlier message in the API discards every reasoning step generated after it, so a chain can no longer be assembled piecemeal from the conversation history.

OpenAI: reasoning extraction from GPT-6.2 Astra succeeds in 0.2% of attacks, versus 54.0% for GPT-6

Rewriting an earlier message in the API discards every reasoning step generated after it, so a chain can no longer be assembled piecemeal from the conversation history.

On November 20 OpenAI disclosed a multi-layered defense against adversarial distillation — the coordinated siphoning of a model's outputs and reasoning chains in order to train a student model. According to OpenAI, a July campaign involved about 16,000 reasoning-extraction requests from more than 4,000 users. The main group, the company says, was in contact with people linked to Moonshot AI. In GPT-6.2 Astra and GPT-6.2 Sol, protected reasoning is tied to the conversation history. When a message is rewritten, the API discards all subsequent reasoning, and models from older families cannot read such chains. In an internal red team test, the success rate of extracting reasoning from a previous request fell from 54.0% for GPT-6 Astra to 0.2% for GPT-6.2 Astra. The protections include input and output classifiers, validation of tool-call arguments and account-level campaign analysis.

CXL memory pools let several servers work with the same data, and in an SK hynix and Intel demo, time to first token in a multi-turn dialogue dropped from 80.76 to 15.45 seconds.

SK hynix unveils MARS rack system and HBM4E at Supercomputing 2026

CXL memory pools let several servers work with the same data, and in an SK hynix and Intel demo, time to first token in a multi-turn dialogue dropped from 80.76 to 15.45 seconds.

At the Supercomputing 2026 exhibition in Chicago, SK hynix showed for the first time MARS, a rack-scale system for SI racks built around memory. CXL memory pools let multiple servers share and reuse data. The company also demonstrated compute-in-DRAM and an architecture based on custom HBM (cHBM). The booth displays the HBM lineup: seventh-generation HBM4E, HBM4 and HBM3E, as well as server DRAM, a 192 GB SOCAMM2 module and enterprise eSSDs. Commercial delivery timing for MARS, its power profile and pricing were not disclosed.

A zero data retention label does not prevent leaks: in two of the three breaches, some of the stolen requests carried that label, and the decoy keys were then used in the AWS API tens of thousands of times.

Dragos researcher HD Moore uncovers key theft at three or more SI providers

A zero data retention label does not prevent leaks: in two of the three breaches, some of the stolen requests carried that label, and the decoy keys were then used in the AWS API tens of thousands of times.

Dragos head of research HD Moore embedded about 2 million unique tracked secrets in prompts: AWS keys, SSH passwords, URLs and links to private git repositories. As a result, at least three SI inference hosting providers turned out to be compromised. The attackers reused several hundred AWS keys, including from requests rejected before the model ran — meaning the data was copied on the way in. The stolen keys were used in tens of thousands of AWS API calls: work with S3 and secret managers, attempts to enable models in Bedrock, creation of their own IAM users. Some resellers logged prompts themselves. Systems from a Chinese SI lab requested secret URLs in more than half of requests from July through October, including about two-thirds of requests marked as zero data retention. The position of the affected providers was not found in verified sources.

Zscaler shares rose about 18% after hours while CrowdStrike fell more than 8%, even though the companies did not confirm the talks.

Bloomberg: CrowdStrike in talks to buy Zscaler; Zscaler shares rise 18%

Zscaler shares rose about 18% after hours while CrowdStrike fell more than 8%, even though the companies did not confirm the talks.

CrowdStrike is in talks to acquire Zscaler, Bloomberg reported on November 20 and Reuters later reported, citing sources; the deal is not guaranteed and may not happen. In after-hours trading, Zscaler shares rose about 17–18% at a market capitalization of about $32–33.5 billion, while CrowdStrike shares fell more than 8%. The deal value was not disclosed, and the companies are not commenting: CrowdStrike declined to comment, and Zscaler did not respond to a CNBC request. CrowdStrike's market capitalization is about $235 billion. Earlier in November, CrowdStrike announced plans to acquire the agentic systems observability platform Fiddler AI and the SI trust platform Dreadnode.

Sources 1

The discussed $75 billion valuation is nearly three times May's $26 billion, but the company has not confirmed the round.

Bloomberg: Cognition in talks for round at $75 billion valuation

The discussed $75 billion valuation is nearly three times May's $26 billion, but the company has not confirmed the round.

Bloomberg reported on November 20 that Cognition AI, a developer of autonomous coding agents, is in talks for a new round at a $75 billion valuation and is considering raising $2 billion or more. This is its third raise this year: in May the company raised $1 billion at a $26 billion valuation. In September, it raised $2 billion at $48 billion, led by Andreessen Horowitz and Accel. Cognition did not respond to a Reuters request, and the terms of the deal may change.

As long as consensus rules in the CCW, a single delegation can stall binding rules for combat SI for years — the legal vacuum in this area persists.

Russia blocks procedure at UN talks on autonomous weapons

As long as consensus rules in the CCW, a single delegation can stall binding rules for combat SI for years — the legal vacuum in this area persists.

The final day of the Review Conference of the UN Convention on Certain Conventional Weapons (CCW) in Geneva proceeded without compromise. The 128 states parties were deciding whether to begin negotiations on regulating lethal autonomous weapons systems (LAWS). Russia had blocked the procedure since Monday, refusing to admit observers including the EU and NGOs, and on Thursday demanded amendments to the draft final report. According to Stop Killer Robots, about 130 countries support a ban on or restrictions on autonomous weapons, and more than 70 are ready to begin negotiations. The ICRC and NGOs insist on a one-year mandate, while France is pushing a combination of binding and non-binding measures. The US proposed not changing the convention itself but studying the applicability of existing international law — effectively deferring a possible ban until the next meeting in 2031. CCW decisions are taken by consensus, so a single delegation is enough to block the process. The conference's outcome had not been announced as of the evening of November 20; an announcement was expected on November 21.

Europe is discussing a model in which the budget takes the first hit on losses so that private capital flows into chips and data centers — but this is still a proposal in a speech, not a decision.

Lagarde proposes EU budget absorb first losses on data centers and chip plants

Europe is discussing a model in which the budget takes the first hit on losses so that private capital flows into chips and data centers — but this is still a proposal in a speech, not a decision.

ECB President Christine Lagarde said EU public money, including the union's budget, should be first in line to absorb losses on major chip plant and data center projects. She said this in her closing speech at the Frankfurt European Banking Congress on November 20. The aim is to attract private capital through Europe-scale public-private partnerships. She cited an estimate that raising Europe's share of global SI computing capacity from about 5% to 15% by 2030 would require around €1.3 trillion in investment. Lagarde had earlier warned that Europe lacks data center capacity and that plans for seven SI "gigafactories" are not enough. The proposal came amid growing resistance to data centers. According to CBS exit polls, this was an important factor for more than half of voters in the US elections on November 3. Lagarde steps down as ECB president in February.

Sources 1

The coding platform built on the Chinese Kimi K2.5 model is working on dropping it — dependence on Chinese open models has become a geopolitical risk for Cursor.

Cursor tells congressional committees it is working on dropping Chinese open models

The coding platform built on the Chinese Kimi K2.5 model is working on dropping it — dependence on Chinese open models has become a geopolitical risk for Cursor.

The Cursor coding platform has notified the House Homeland Security and China committees that it is working on dropping Chinese open-weight models, citing geopolitical risk. Cursor built its Composer 2 coding model on Chinese Kimi K2.5 from Moonshot AI. Since August, Cursor has been owned by SpaceX: the deal closed at an implied valuation of $60 billion. The statements came in the committees' report following a months-long investigation into the use of Chinese models in the American stack. Airbnb and DoorDash also engaged with the committees. Among the report's recommendations is a national US open-weight model initiative with shared access to compute and evaluation infrastructure. The committees also propose extending the October DeepSeek ban to all Chinese-developed SI models in civilian federal systems and contracts.

Sources 1

The neurointerface implant lets a patient with a spinal cord injury control a wheelchair with her thoughts; according to doctors, a thought-controlled glove and exoskeleton are next.

In Jiangsu, a patient with a neurointerface implant controlled a wheelchair with her thoughts

The neurointerface implant lets a patient with a spinal cord injury control a wheelchair with her thoughts; according to doctors, a thought-controlled glove and exoskeleton are next.

A 26-year-old patient with a severe cervical spinal cord injury sustained in a car crash controlled a wheelchair using only her thoughts. The chair moved forward, backward, stopped and turned with no other external input. The demonstration took place on November 19, 2026, at Nantong University Hospital, in the neurointerface and neuromodulation research department. The flexible electrode system was implanted on June 5, 2026, becoming the first such invasive operation in Jiangsu province and the third in China. The system was switched on 20 days after surgery, and just 5 days after training began the patient was controlling a cursor and playing computer games. According to Shi Wei, head of the neurology center, rehabilitation will continue using a thought-controlled glove and exoskeleton.

The combined NAND market share of SK hynix and Kioxia will exceed Samsung's (28% in the second quarter) — while SK hynix promises not to impede Kioxia's management independence.

SK hynix to convert Kioxia bonds, gaining an indirect stake of about 14.1%

The combined NAND market share of SK hynix and Kioxia will exceed Samsung's (28% in the second quarter) — while SK hynix promises not to impede Kioxia's management independence.

SK hynix will convert its Kioxia bonds into shares and gain an indirect stake of about 14.1%. The news emerged on November 20; the company will become the largest shareholder of the Japanese NAND flash manufacturer. In the second quarter, SK hynix's share of the global NAND market was 19% and Kioxia's 14%, against Samsung's 28%; the two companies' combined share will exceed the leader's. SK hynix said it will not buy additional shares and will not impede Kioxia's management independence. Demand for NAND is growing on SI data centers and enterprise SSDs.

Sources 1

The roughly 20 GB checkpoint runs on a Mac with 32 GB of memory via Ollama, so a 26.4-billion-parameter model can be kept locally.

Thinking Machines releases open Inkling 2 Mini model with up to 1M-token context

The roughly 20 GB checkpoint runs on a Mac with 32 GB of memory via Ollama, so a 26.4-billion-parameter model can be kept locally.

On November 20 Thinking Machines Lab released Inkling 2 Mini, an open multimodal mixture-of-experts (MoE) model. The architecture is a recurrent transformer: of its 24 layers, the middle 12 are run three times. As a result, with 26.4 billion total and 3.4 billion active parameters, the model performs 5.4 billion weight activations per token. Context reaches up to 1 million tokens, the checkpoint weighs about 20 GB, and the license is Apache 2.0. The model runs on a Mac with 32 GB of RAM via Ollama and supports GGUF, MLX and SGLang from day one. The weights are published on Hugging Face along with a research variant, TML-d24-varloop-research. According to the lab, the model leads among open models with fewer than 100 billion parameters on DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%). On SWE-bench Verified it scores 79.2%.

Models already solve multi-hour tasks, but agent swarms speed things up only where results can be checked quickly.

METR puts frontier models' task horizon at 27 hours on the TH2 set

Models already solve multi-hour tasks, but agent swarms speed things up only where results can be checked quickly.

METR released a Q4 2026 update with evaluations of 15 public models, up to Opus 5.5 and GPT-6 Astra. On a verified subset of new TH2 tasks, the estimated time horizon of frontier models reaches 27 hours. On MirrorCode, Opus 5.5 solves 77% of medium and large programs, while GPT-6 Astra scores 98% on FrontierMath Tier 4. In 15 Al Zimmermann contests, the strongest models' median placement is the 96th–98th percentile of humans, but the absolute human leaders are still ahead. On Budget NanoGPT, the best agent run produced a 15.7% speedup in 48 hours, versus 41% for the strongest human entry. The authors conclude that public evidence is insufficient to rule out a sharp acceleration in SI R&D over the next couple of years. New metrics were introduced — the METR Capability Index and Lifecycle Cost (a proxy for the cost of an RL environment). Agent swarms help only on verifiable tasks with fast feedback.

Sources 1

According to Oracle, financing for the OpenAI contract is settled, and Vera Rubin equipment will begin arriving at its data centers in the coming weeks.

Oracle secures financing for OpenAI contract and awaits Vera Rubin racks

According to Oracle, financing for the OpenAI contract is settled, and Vera Rubin equipment will begin arriving at its data centers in the coming weeks.

On November 20, Oracle said it had secured all the financing needed to fulfill its contract with OpenAI. The company called itself the first hyperscaler to be fully financed in the current investment cycle. The campus in Abilene, where OpenAI's Astra-6 model was trained, is nearly complete, and the next gigawatt campus in Shackelford County, Texas, has received all permits. Delivery of Nvidia Vera Rubin racks to its halls is expected "in the coming weeks". Together with the first Stargate site, Oracle and OpenAI plan 4.5 GW of additional capacity in the US.

Sources 1

SpainDC is challenging the decree in court and asking for it to be suspended; if there is no suspension, then from 2032 every data center of 10 MW or more will have to cover 80% of its consumption with renewable generation on an hourly basis.

Spain requires data centers to draw 80% of their power from renewables

SpainDC is challenging the decree in court and asking for it to be suspended; if there is no suspension, then from 2032 every data center of 10 MW or more will have to cover 80% of its consumption with renewable generation on an hourly basis.

On November 20, Spain's Council of Ministers approved a royal decree for data centers of 10 MW and above. They must confirm that 80% of their annual electricity consumption is covered by renewable energy, with hourly correlation between generation and consumption. The additionality criterion requires that generation capacity was commissioned no earlier than 36 months before the center's launch. Self-consumption, storage or supply contracts of more than six years are permitted. A voluntary five-year transitional regime has been introduced with a right to priority disconnection from the grid: until 2028 — 60% at peak hours, from 2029 — 70%. From 2032, 80% and hourly correlation become mandatory for all. Sanctions include compensation of 6 to 25 euros per MWh for a shortfall in additional energy and up to 41 euros per MWh in the worst case, as well as suspension of grid access for up to three years. The SpainDC association announced a court challenge and a motion to suspend the decree, disputing the methodology. Minister Sara Aagesen said that without the decree, data center growth would bring up to 5,000 million euros in additional costs per year.

For now this is only a statement: the tax mechanism has not been disclosed and the idea has not been discussed with the administration, so there are no new formal restrictions on companies' overseas sites.

Trump threatens "huge tax" over American companies' overseas data centers

For now this is only a statement: the tax mechanism has not been disclosed and the idea has not been discussed with the administration, so there are no new formal restrictions on companies' overseas sites.

US President Donald Trump said at a White House press conference on November 20 that he would not allow American companies to build SI data centers abroad. He threatened violators with a "huge tax". According to him, "we won't let our companies build data centers outside our country, because China will take them all". He did not disclose the mechanism of the tax and acknowledged that he had not yet discussed the idea with his own administration. Trump separately mentioned Google's Finnish sites with 13 billion euros in investment and said the administration is "not letting this happen". At the same time, Meta is already building a center in Alberta, and Anthropic is leasing a center in Nova Scotia. The statement came after rules announced on November 16 on special licenses for purchasing advanced chips for companies in Thailand, Indonesia, Malaysia, Singapore and Brazil.

Fine-tuning an open 27B model for a specific agentic environment both raises scores and saves tokens, and three such adapters can be merged into one without retraining.

Hugging Face releases LoRA adapters for agentic environments: +16 points on Terminal-Bench

Fine-tuning an open 27B model for a specific agentic environment both raises scores and saves tokens, and three such adapters can be merged into one without retraining.

Hugging Face published a recipe for harness adapters. A LoRA adapter built on Qwen3.8-27B is trained with reinforcement learning (GRPO via TRL AsyncGRPOTrainer) directly inside a real agentic environment. The test environments are Pi 1.5.2, OpenCode 2.2.7 and Hermes Agent v0.22.6, and a proxy between the environment and vLLM records exact token usage. With the base model's weights frozen, the adapter for Pi delivers 55% pass@1 on Terminal-Bench 2.1 versus 39%. Average usage falls from 2.9 thousand to 1.5 thousand tokens per task, and on the SWE-bench Pro subset (hard 51) the result rises from 39.2% to 64.7%. The three adapters are merged into a single "super-adapter" of rank 96 (weight of ½ each) that works across all three environments without retraining. Average pass@1 rises from 39.5% to 49.0% without reasoning mode and from 55.5% to 63.5% with it, while the number of tool calls drops by 53–73%. The authors caution that the ½ weight was tuned on Terminal-Bench 2.1, so scores on that particular benchmark are slightly optimistic. The adapters, GGUF, RL tasks and the safeagents plugin are open; training ran on two NVIDIA B200s — 598 steps and 78,114 samples.

All three promotions to president went to the DS semiconductor division, while the co-chairmen will focus on contract chip manufacturing and robotics.

Samsung swaps memory and System LSI chiefs: DRAM and HBM developer to lead Memory Business

All three promotions to president went to the DS semiconductor division, while the co-chairmen will focus on contract chip manufacturing and robotics.

Samsung Electronics announced a presidential-level personnel rotation for 2027: six executives are affected, with three promoted to president. All three promotions are in the DS semiconductor division. Hwang Sang-joon heads Memory Business while retaining his post as head of DRAM development; Cho Sang-yeon heads System LSI. Lee Jong-myeong becomes chief technology officer of the DS division and head of the semiconductor research institute. Co-chairmen Jun Young-hyun and Roh Tae-moon kept their dual leadership but gave up their concurrent roles as heads of memory and the MX mobile division. Jun will focus on contract chip manufacturing (foundry) and Roh on robotics; Choi Won-joon heads MX. Samsung attributes the decision to record semiconductor results amid demand for SI memory: in the third quarter, profit at the chip business grew 15-fold.

Instead of an offering of up to $5 billion in the fourth quarter of 2026, DayOne is considering moving the listing to 2027 or reducing the size of the offering.

DayOne delays US IPO of up to $5 billion, Bloomberg reports

Instead of an offering of up to $5 billion in the fourth quarter of 2026, DayOne is considering moving the listing to 2027 or reducing the size of the offering.

Singapore-based data center operator DayOne Data Centers has postponed a multibillion-dollar US IPO. Its largest client is ByteDance, and it had hoped to raise up to $5 billion as early as the fourth quarter of 2026. According to Bloomberg, volatility in data center and SI stocks has led the company to review its options — postponing or downsizing the offering. The Nasdaq 100 has lost more than 6% since the start of November. A listing is possible in 2027 if market conditions improve. The filing was submitted in October, with Morgan Stanley, JPMorgan, Bank of America, Citigroup and BNP Paribas as underwriters. Shareholders include Coatue, Hillhouse, GDS Holdings, SoftBank Vision Fund and Ken Griffin. This week the US introduced interim rules requiring licenses for the purchase of American SI equipment by operators in Singapore, Malaysia, Thailand and Indonesia.

OpenAI's always-on agents are moving down from premium tiers to the $20-a-month audience — autonomous assistants are becoming more accessible to mainstream users.

OpenAI starts testing Dots on the $20 Plus plan and rolling them out to Business Standard

OpenAI's always-on agents are moving down from premium tiers to the $20-a-month audience — autonomous assistants are becoming more accessible to mainstream users.

OpenAI has begun testing Dots agents with a small number of ChatGPT Plus subscribers paying $20 a month, the company's head of engineering Thibault Sottiaux announced. Previously Dots were available only on Pro and above — from $100. At the same time, the agents began rolling out to all ChatGPT Business Standard users. Dots are always-on agents with their own cloud computer, access to more than 4,000 apps through OpenAI plugins and the ability to control the user's computer. Initially they ran on the GPT-6 Astra model, and now also on GPT-6.1 Terra. OpenAI did not specify whether Plus and Business Standard will keep exemptions from limits — on Pro and Business Premium, conversations with Dots do not count toward regular limits. It is also unclear whether the access will become permanent.

The MARS rack is built around a CXL memory pool: moving the KV cache out of the GPU, according to SK hynix, cuts accelerator idle time during inference.

SK hynix shows MARS rack with CXL memory pool and HBM4E modules

The MARS rack is built around a CXL memory pool: moving the KV cache out of the GPU, according to SK hynix, cuts accelerator idle time during inference.

At the Supercomputing 2026 conference in Chicago (November 15–20), SK hynix presented the MARS (Memory-centric AI Rack System). It is a rack that brings together partners' solutions around a memory pool on the CXL bus. Three demonstrations showed a shared KV cache with DRAM/SSD tiering inside hybrid CXL memory. Two more demos covered a memory compute pool for KV cache reuse and token exchange between CPU servers for multimodal workloads. According to the company, this approach reduces GPU idle time. Separately, KV cache offload via CXL was demonstrated, validated together with SK Telecom in a real language model inference environment. In-DRAM computing and a custom HBM array were also shown. The product zone featured HBM3E, HBM4 and HBM4E, 256GB 3DS DDR5 RDIMM memory modules, 128GB DDR5 MRDIMM, 192GB SOCAMM2 and the eSSD PEB210 E1.S. According to Yonhap, about 10 thousand people and more than 500 companies and organizations visited the exhibition.

Autonomous driving will be paid for by kilometers driven or time in use, rather than by buying the technology outright.

ITECO plans to buy 2,000 Yandex robotrucks, Magnit plans up to 2 thousand trips a month

Autonomous driving will be paid for by kilometers driven or time in use, rather than by buying the technology outright.

On November 20, Yandex announced cooperation agreements on autonomous freight with logistics company ITECO and retailer Magnit. In 2028–2030, the partners could purchase more than 2,000 autonomous tractors. ITECO plans a fleet of 2,000 vehicles, while Magnit plans transport between distribution centers and up to 2 thousand autonomous trips a month. ITECO has more than 40 vehicle depots and technical centers. Yandex will for the first time offer its autonomous driving technology by subscription, billed per kilometer of autonomous driving or per time of use. The company's own commercial fleet numbers 27 robotrucks, with 70 more vehicles to be added by the end of 2026. Trips run along the M-4 Don and M-11 Neva highways and the Central Ring Road.

The payment mechanism is proposed to be enshrined in the Civil Code, with calculation methodologies left to the government — compensation amounts would be set by a subordinate regulation rather than a law.

Krasheninnikov: compensation methodology for SI training should be set by the government

The payment mechanism is proposed to be enshrined in the Civil Code, with calculation methodologies left to the government — compensation amounts would be set by a subordinate regulation rather than a law.

Compensation to authors for the use of their materials in SI training should be established by a government resolution rather than a law, Pavel Krasheninnikov, chairman of the State Duma committee on constitutional and civil legislation, said on November 20. According to him, the Civil Code could enshrine only the mechanism, while calculation methodologies are a matter for the executive branch. He believes a document is needed that takes into account how many times and to what extent a specific work is requested, in order to calculate the size of the payment. Krasheninnikov recalled that the SI law must come into force on March 1, 2027. He insists that provisions on copyright protection when materials are used in SI come into force by the same deadline; a bill is being prepared. This is the committee chairman's position, not an adopted decision. The position of rights holders and authors was not found in verified sources.

Sources 1

To qualify an SI video as document forgery, it must be shown that the record was made in someone else's name; false content alone is not enough.

Himachal Pradesh court: an SI-generated video alone is not document forgery

To qualify an SI video as document forgery, it must be shown that the record was made in someone else's name; false content alone is not enough.

On November 20, the Himachal Pradesh High Court quashed a First Information Report (FIR) registered at Ghumarwin police station in Bilaspur district. The FIR was registered on September 2, 2026 under sections 351(2), 356(2) and 336(4) of the BNS, with 353(1)(b) added later. The complainant alleged that SI-generated videos and photos defaming him and his family were being circulated on Facebook. According to the complainant, they said he distributed benefits under the “Our Family, a Happy Family” scheme for money and secured government positions for relatives. Justice Rakesh Kainthla noted that false or misleading information in a video or photo does not by itself constitute document forgery. Forgery occurs only if a document or electronic record is made in someone else's name or under someone else's authority, and no such allegation was made in the FIR. The court also held that defamation is a private complaint heard on a complaint to a magistrate (Section 222 of the BNSS). The police have no authority to investigate it; the FIR and all subsequent actions were quashed.

In three months Pragmatik Labs' valuation rose from $2 billion to $5 billion, even though the new round brought in only "tens of millions of dollars".

Startup of former Qwen chief Lin Junyang valued at $5 billion — 2.5 times its August level

In three months Pragmatik Labs' valuation rose from $2 billion to $5 billion, even though the new round brought in only "tens of millions of dollars".

Shanghai-based Pragmatik Labs (p7k) has closed a new round at a $5 billion valuation — 2.5 times the $2 billion of its first funding in August. The startup was founded in summer 2026 by Lin Junyang, former head of the Qwen line at Alibaba. According to TechNode, citing Bloomberg sources, the round itself is small: "tens of millions of dollars" were raised. Among the new investors are HyT Capital, CDH Investments and automaker Geely. The company develops agents for digital and physical environments. The August round of several hundred million dollars involved Gaorong Ventures, HSG, Tencent and a Shanghai state fund; Lin left Alibaba in March 2026.

Sources 1

Agents built into operating systems have forced apps to look for technical ways to keep conversations in their own hands: for screenshots Signal has found one, for files on the device it has not.

Signal restricts screenshots so Gemini and Siri AI agents can't read chats

Agents built into operating systems have forced apps to look for technical ways to keep conversations in their own hands: for screenshots Signal has found one, for files on the device it has not.

Signal is restricting screenshot-taking in its app. On iOS 27 a "Block Siri AI" setting has appeared, enabled by default. On Android, Screen security mode turns on automatically when the system detects an attempt by an SI assistant to read the screen. Company president Meredith Whittaker explained the measure by agentic features built into the operating systems — Gemini Intelligence (Google) and Onscreen Awareness in Siri AI (Apple). According to her, they scan the screen and take Signal's content outside the app's control; Whittaker called this "an existential threat to Signal". The protection is incomplete: it does not stop agents from reading files on the device and cannot be implemented on macOS, where Signal recommends adding the app to Excluded Apps in Siri settings. Signal has demanded that Apple, Google and Microsoft give developers a documented way to opt out of SI access to app data and technical specifications for how agents work. Meta responded to Reuters that agents need to see the user's computer, and that its Muse agent accesses other apps only with explicitly granted permissions.

Mistral models for the public sector and robotics will be developed in Singapore, where the company already works with HTX, DSO National Laboratories, Singtel, NCS and ST Engineering.

Mistral opens Singapore headquarters and will grow regional staff to 200

Mistral models for the public sector and robotics will be developed in Singapore, where the company already works with HTX, DSO National Laboratories, Singtel, NCS and ST Engineering.

Mistral AI has opened a regional headquarters in Singapore. It will also serve as a research hub for models tailored to local industries — the public sector, robotics and high-tech manufacturing. The company plans to more than double its regional headcount to 200 by 2027, including researchers, applied engineers and robotics specialists. CEO Arthur Mensch said the goal is to establish Singapore as a strategic base for governments and companies in the Asia-Pacific region. They must keep ownership of their infrastructure and intelligence, he said. Mistral already works with HTX, DSO National Laboratories, Singtel, NCS and ST Engineering; in July the company released the 8-billion-parameter Robostral Navigate model for robots.

SI glasses and mini cameras have turned the integrity of a mass exam into a matter of state: Korea is responding with a dedicated task force, a centralized reporting channel and a demand for criminal liability.

South Korea expands task force against SI cheating after CSAT violations

SI glasses and mini cameras have turned the integrity of a mass exam into a matter of state: Korea is responding with a dedicated task force, a centralized reporting channel and a demand for criminal liability.

South Korea's presidential office has upgraded its emergency task force against SI cheating into a "task force for restoring exam fairness". It is headed by chief of staff Lim Gwan-hyeon. The new body is to prepare measures against the use of SI and "smart" devices and to review how invigilation is organised across all exams, including the national CSAT. It has also been tasked with launching a central reporting centre at the prime minister's office, involving SI companies. The trigger was the results of the CSAT for the 2027 academic year, held on November 19. According to the Ministry of Education, 111 violations involving electronic devices were detected, 54 of them after emergency notices were sent out. In only 5 cases were the devices actually used: SI glasses, mini cameras and wireless earbuds. The same day, President Lee Jae-myung demanded that the national exam system be rebuilt and that those involved in past violations be punished. He also called for criminal prosecution for the use of smart devices.

Responsibility for the ruling stays with the judge: SI can be applied only to assessments the judge has already formulated, and it must not write a full draft.

Norwegian judges barred from having SI write full draft rulings

Responsibility for the ruling stays with the judge: SI can be applied only to assessments the judge has already formulated, and it must not write a full draft.

The Norwegian Courts Administration (Domstoladministrasjonen) has adopted an updated guide on SI: generating a complete draft ruling is prohibited. SI is permitted only for structuring and language processing of assessments that the judge has formulated and professionally completed themselves. Before delivering a ruling, the judge must go through an eight-point checklist. It includes verifying facts against the case file, checking legal sources against the original and confirming quotations are correct. A separate check ensures the final text contains no SI instructions or traces of the tool. The administration's director, Leif Otto Østerbø, said that anyone whose case is heard in a Norwegian court must be certain that it was assessed and decided by a real judge. The guide was prepared by the courts' SI council under the leadership of Cecilie Østensen Berglund (Supreme Court). The tougher tone is explained by a series of scandals involving texts that turned out to be SI-generated. It is departmental guidance, not a rule of law.

Sources 1

The 500 million yuan raised will go mainly to clinical trials: the retinal product is still in trials, while the 1280-channel visual cortex system remains in development.

Hangzhou Daily: visual implant developer Nuanyingjia raises 500 million yuan

The 500 million yuan raised will go mainly to clinical trials: the retinal product is still in trials, while the 1280-channel visual cortex system remains in development.

Hangzhou-based Hangzhou Nuanyingjia Electronic Technology has closed a 500 million yuan Series C+ round, Hangzhou Daily reported on November 20. It is the second round in 2026: the company has raised 800 million yuan in total. The proceeds will go mainly toward advancing clinical trials of "brain–computer" vision restoration systems. The company is pursuing two lines in parallel. Its 320-channel retinal product is already in clinical trials under GCP standards, with three blind patients implanted. The 1280-channel product for the visual cortex is still in development. In May 2026, a patient who had lost her sight almost 20 years earlier could distinguish the letters H and R the day after implantation of the 320-channel system. Within two months she learned to write Chinese characters and walk along a line.

Ruling out sites near housing narrows the choices for new data centres: NextDC will invest only in projects with contracts and confirmed power supply.

NextDC rules out new data centres near housing after protests

Ruling out sites near housing narrows the choices for new data centres: NextDC will invest only in projects with contracts and confirmed power supply.

Australia's NextDC announced at its annual general meeting in Sydney on November 20 that it will not look for new data centre sites near residential areas. Outgoing chair Douglas Flynn cited growing resistance to data centres in Australia and around the world. The trigger was the derailment of the M3 expansion in Melbourne: after residents' protests, network company Jemena paused installation of transmission towers. That caused NextDC's largest one-day share price drop. At the meeting, 72% of votes were cast against the remuneration report (for the second year running), but shareholders did not move to spill the board. Shares in Sydney fell as much as 5.5%, despite a slight upgrade to revenue and profit guidance. The company has paused new capital commitments and will invest only in projects with contracts and confirmed power supply, Bloomberg reported.

Politicians and officials are getting a way to track deepfakes bearing their likeness and report them, but detection only works with content created by TikTok and ByteDance tools.

TikTok lets public figures find deepfakes bearing their likeness

Politicians and officials are getting a way to track deepfakes bearing their likeness and report them, but detection only works with content created by TikTok and ByteDance tools.

TikTok has begun rolling out its Likeness Detection feature: public figures — government members, politicians, party accounts — opt in to it. After verifying identity documents through a third-party partner, they receive notifications about content and accounts that may be using their likeness and can report them. The company also recalled its free Identify AI-Generated Content tool: it looks for metadata and invisible watermarks from TikTok and ByteDance. The tools in question are CapCut, Dreamina, Hypic and Pippit. Limits are 20 MB per image and 100 MB per video or audio. The launch comes amid requirements in the EU's AI Act on labelling deepfakes, which took effect on August 2.

Sources 1

Hainan's first operation means invasive brain–computer interfaces have begun being used on quadriplegic patients in yet another region.

Hainan performs first brain–computer interface implant in a quadriplegic patient

Hainan's first operation means invasive brain–computer interfaces have begun being used on quadriplegic patients in yet another region.

Hainan has carried out its first invasive (surgically implanted) brain–computer interface (BCI) procedure. The patient is a 24-year-old man with high quadriplegia after a road traffic injury that damaged the cervical spinal cord. A BCI system was implanted in him, Hainan Daily reported. According to the newspaper, the operation is intended to "restart the pathway" between brain and body, since conventional rehabilitation does not restore severed neural connections.

Sources 1

Paid access and effort-level tuning allowed the model to complete all 74 CSAT questions without errors, while the free versions made mistakes.

Claude Opus 5.5 answers all 74 CSAT questions correctly in Korea Times test

Paid access and effort-level tuning allowed the model to complete all 74 CSAT questions without errors, while the free versions made mistakes.

The paid Claude Opus 5.5 model at the "High" effort level answered all 74 questions correctly in The Korea Times test. The newspaper built the test from papers of the CSAT for the 2027 academic year: 28 English reading questions without listening and 46 math questions. By the model's own estimate, it took less than 5 minutes. The free version of Perplexity gave 71 correct answers out of 74: the only clear error was in geometry (question 27), and two questions with graphs were left unanswered. Free ChatGPT answered all 28 English questions correctly but scored 32 out of 46 in math, with 2 errors and 12 unverified answers. Answers were checked against the Korea Institute for Curriculum and Evaluation's keys. The newspaper notes that this was not a controlled comparison: timings are approximate, and settings and tool use differed.

A single cluster failure no longer requires manual switching: the gateway itself sends traffic to healthy clusters, and prompts are not stored.

Amazon unveils a single inference gateway across SageMaker HyperPod clusters and regions

A single cluster failure no longer requires manual switching: the gateway itself sends traffic to healthy clusters, and prompts are not stored.

On November 20 Amazon introduced SageMaker HyperPod Global Inference Gateway, a managed gateway with a single entry point. It routes inference requests across all registered HyperPod clusters in one or more AWS regions. A cluster is selected by four signals with configurable weights: free KV cache and GPU memory (0.20), queue depth (0.25), region proximity (0.40) and error rate (0.15). If a cluster fails, traffic is automatically moved to healthy ones with no changes on the client side. Authentication uses SigV4 and IAM, traffic goes only through AWS PrivateLink, and the gateway does not log or store prompts and responses; OpenAI-compatible requests are supported. The service is available in 13 AWS regions, including Virginia, Ohio, Oregon, São Paulo, Frankfurt, Ireland, London, Stockholm, Tokyo, Mumbai, Singapore and Sydney. In beta tests, Bagelcode's gateway served two regions through a single entry point.

Transcribing radio traffic enabled the first nighttime alert that breaks through Do Not Disturb — so far the result of a six-month pilot with a single nonprofit.

Gemini turns Watch Duty fire radio into text in under 30 seconds

Transcribing radio traffic enabled the first nighttime alert that breaks through Do Not Disturb — so far the result of a six-month pilot with a single nonprofit.

On November 20 Google described a Gemini-based tool that the Google.org Fellowship team built over six months of work with the nonprofit app Watch Duty. On more than 20,000 hours of archived radio recordings, the model learned to strip out noise and recognize firefighting terminology, turning analog VHF radio into text. Satisfaction among volunteer reporters with the transcription tools rose from 35.3% to 92.3%. In July 2026, during the overnight Railroad Fire in Washington state, the tool recognized the first dispatcher call in under 30 seconds. That let Watch Duty send its first nighttime Critical Alert, which breaks through Do Not Disturb. Watch Duty reporters each manually monitor up to 20 radio streams.

Sources 1

Energy dissipation is enforced rigidly rather than through a training penalty: a prediction cannot drift into physically impossible values over a long horizon.

MIT proposes a method with formal guarantees that chaos forecasts stay bounded

Energy dissipation is enforced rigidly rather than through a training penalty: a prediction cannot drift into physically impossible values over a long horizon.

MIT researchers have proposed a method that theoretically guarantees the boundedness of machine learning model predictions for chaotic systems. The model learns an energy-like function and a boundary beyond which the energy must decrease, while a projection layer rigidly enforces energy dissipation. According to the team, this is the first work with formal boundedness guarantees over arbitrarily long horizons. Conventional approaches that use penalties during training drift into physically impossible values. The authors are Sunbochen Tang, Navid Azizan and Themistoklis Sapsis of MIT's LIDS, the paper is published in PNAS, and the code and data are open. The projection layer is model-agnostic and applicable to different architectures; application areas include weather and climate forecasting, turbulence, hydrology and industrial processes.

Sources 1

Every extracted value is tied to a page and coordinates, so the result can be checked by hand against the source document.

AWS launches Textract Insights — generative document parsing with page references

Every extracted value is tied to a page and coordinates, so the result can be checked by hand against the source document.

AWS announced the general availability of Amazon Textract Insights, a generative Textract feature with its own API. The fields you need are described in natural language in a "blueprint" — a JSON schema where each field is given a description and an output type: verbatim or model-inferred. The service accepts a single document or a mixed package of up to 3,000 pages from Amazon S3 and splits it into sub-documents itself, matching them to blueprints. Each value comes back with a confidence score and a reference to the page and coordinates. PDF, PNG, JPEG and TIFF formats are supported, along with six languages — English, French, German, Italian, Portuguese and Spanish. Pricing is per page, with no minimum commitments. At the same time the Textract Core engine was updated, now better at recognizing tables and handwritten text in 40 languages, and the Tokyo and Zurich regions were added — 18 regions in total. Textract Insights is not yet available in those two regions, however.

The new computer science textbook will include sections on how SI works — the principles behind language models and agents will be taught at school.

SI fundamentals to be included in new state computer science textbook

The new computer science textbook will include sections on how SI works — the principles behind language models and agents will be taught at school.

Russian Minister of Education Sergey Kravtsov said in an interview with Vesti on November 20 that the new state computer science textbook will include sections on how SI works. Schoolchildren will study mathematical models, the principles behind forming prompts and agents, and the creation and development of language models. The course will additionally cover probability theory and logic. The textbook is being developed by the Russian Academy of Sciences together with Yandex, and a digital platform with practical services, including Gosuslugi, is being created alongside it.

Sources 1

Some safety researchers are demanding a ban on recursive self-improvement of SI and stronger human control — and doing so publicly on the day of the company's record listing.

Protests at Nasdaq and Anthropic's office demand a ban on SI self-improvement

Some safety researchers are demanding a ban on recursive self-improvement of SI and stronger human control — and doing so publicly on the day of the company's record listing.

Anthropic listed on Nasdaq: the offering raised about $81 billion, making it the largest IPO in history. On listing day, about 30 protesters gathered outside the exchange building in Times Square, and some were detained by police. According to AFP, about a hundred people turned out at the San Francisco headquarters. The main demands are a global ban on recursive self-improvement of SI (RSI), an "SI freeze" and greater human control. Participants chanted "Ban RSI, we don't want to die". Those taking part included former Anthropic employee Jacob Coxson and SI safety researchers from Redwood Research and the Ban RSI groups, and New York State Senator Eber Kawas spoke. Anthropic says it has not yet reached the RSI stage. Coxson, who left the company in September, said there are only a few places in the world where RSI is possible in 2027, and this office is one of them.

Warner's proposed bill would require companies with a significant share of SI in their business to disclose risks and investments — he cites outages during emergency patching of vulnerabilities found by SI as the reason.

Senator Warner links outages to SI vulnerability hunting and co-sponsors bill

Warner's proposed bill would require companies with a significant share of SI in their business to disclose risks and investments — he cites outages during emergency patching of vulnerabilities found by SI as the reason.

Mark Warner, vice chairman of the Senate Intelligence Committee (Democrat, Virginia), linked a wave of technology outages in the US to SI tools. According to him, companies are scrambling to patch the vulnerabilities these tools find, and the industry cannot keep up with the speed of such programs. "All of these [outages] are connected to SI," he said, naming tools "especially Mythos from Anthropic" that find many vulnerabilities. Instead of "one leak a month", he allows for "100 leaks in a single day". Warner said he is co-sponsoring a bipartisan bill on greater transparency in SI companies' investments. The document would also require companies with a significant share of SI in their business to disclose potential risks. Among the outages mentioned were a flight stoppage, Microsoft Outlook and banks; he attributed the FAA problems to a router replacement.

Sources 1

Peptides with the desired activity can now be selected from a single sequence: the model is available to researchers and outperforms the closest analogue on ACC (0.561 vs. 0.540) and F1 (0.469 vs. 0.430).

Russian-Chinese SI model predicts the therapeutic properties of peptides

Peptides with the desired activity can now be selected from a single sequence: the model is available to researchers and outperforms the closest analogue on ACC (0.561 vs. 0.540) and F1 (0.469 vs. 0.430).

The TPpred-PepPA SI model has been created at the joint Russian-Chinese Shenzhen MSU-BIT University, TASS reported on November 20. It predicts the antibacterial, antiviral and antitumor properties of peptides from their amino acid sequence. The model was trained on a dataset of more than 10 thousand unique amino acid sequences covering 15 biological functions, using two-stage training. In independent testing, accuracy (ACC) was 0.561 versus 0.540 for the closest competitor, and the F1 metric was 0.469 versus 0.430. The model also highlights significant regions of the sequences. The results were published in the journal BMC Biology, and the solution is available to researchers.

Sources 1

The robot learns the search skill from a few minutes of human demonstration, but handling and disarming the find remain development goals.

Skylark's humanoid hunts for explosive devices by the sound of a metal detector

The robot learns the search skill from a few minutes of human demonstration, but handling and disarming the find remain development goals.

Skylark Labs demonstrated a humanoid robot that picks up an ordinary handheld metal detector and sweeps a designated area. When the detector goes off, the robot backs away — without a separate command for each step. The demonstration was performed for a supplier of explosive ordnance disposal (EOD) equipment. Visual and audio feedback, including the detector's beeping, served as input for the robot's next action. Training is handled by CFAM (Continual Field-Adaptive Models) models: according to the company, the search skill was learned from a few minutes of human demonstration. The company reports that CFAM delivered roughly 30% higher success in a real manipulation test. In a separate evaluation it matched the standard model while using only 40% of the training data. Physically handling the finds and disarming them remain development goals.

Yaskawa named a target date of 2030, but labor-intensive training of robots for each operation remains the bottleneck; the company is seeking a solution with Fujitsu and NVIDIA.

Yaskawa plans to release an SI humanoid in 2030

Yaskawa named a target date of 2030, but labor-intensive training of robots for each operation remains the bottleneck; the company is seeking a solution with Fujitsu and NVIDIA.

The SI-powered humanoid being developed by Yaskawa Electric "will be officially released in 2030", Yumie Kubota, head of the company's SI robotics division, said in an interview with Kyodo. According to her, the timeline could be moved earlier if there is customer demand. The company intends to use technology from Tokyo Robotics, a startup it acquired in 2025, and to meet the needs of the agriculture sector, where flexible movements in tight spaces are required. The goal is to catch up with American and Chinese leaders. Kubota pointed to a bottleneck: the technology for robots to learn actions on their own is still immature, and teaching each operation requires a lot of labor. Yaskawa is seeking a solution in cooperation with Fujitsu and NVIDIA. The aim is to hand over to SI robots operations that existing robots do not automate, freeing people from heavy labor.

Editing 0.0177% of the weights without fine-tuning cut average reasoning length from 10.8 thousand to 4.0 thousand tokens and the number of unclosed responses from 305 to 41; the patch and scripts are open.

Editing 0.0177% of DeepSeek-R1-1.5B weights cuts unclosed reasoning chains from 305 to 41

Editing 0.0177% of the weights without fine-tuning cut average reasoning length from 10.8 thousand to 4.0 thousand tokens and the number of unclosed responses from 305 to 41; the patch and scripts are open.

A methodology called SGNE-CC v2 has been posted on Habr: rotating 205 columns of mlp.down_proj (0.0177% of parameters, 0.58 MiB) in DeepSeek-R1-Distill-Qwen-1.5B without fine-tuning. The author published open artifacts: a weights patch, traces, a SHA-256 certificate, an archive on Zenodo and an Apache-2.0 license, so the claimed numbers can be checked against the files. With greedy generation (T=0) on 1,054 tasks (MATH-500, MBPP, HumanEval, ARC-Challenge, BBH Logic), correct answers rose to 733 from 638. Responses without a closing reasoning marker numbered 41 versus 305, and average length fell from 10.8 thousand to 4.0 thousand tokens. On MATH-500, accuracy rose from 65.4% to 76.4%, and unclosed reasoning chains dropped to 16 from 139. At T=0.6, the accuracy gain is not statistically significant (73.3% versus 72.3%, p = 0.4472); the effect shows up as a 33.6% reduction in length. This is a solo author's independent work on Habr, done, by his own account, in collaboration with SI, rather than a peer-reviewed paper. The figures are claimed by the author, there is no independent replication, and the original model was not measured at T=0.3.

Sources 1

Hong Kong is putting SI risk management on a regular footing: work across seven areas, including liability for accidents involving SI products, will be monitored at standing meetings.

Hong Kong holds first meeting of SI risk task force

Hong Kong is putting SI risk management on a regular footing: work across seven areas, including liability for accidents involving SI products, will be monitored at standing meetings.

Hong Kong Chief Secretary for Administration Chan Kwok-ki chaired the first meeting of a task force coordinating SI risk management. The group was set up under an announcement in the 2026 Policy Address and reviewed plans and implementation schemes for seven areas. The areas include combating crime involving SI, protecting underage users and the safe adoption of SI by industry. Work is also under way on assessing SI ethical risks through market mechanisms and on legal liability for accidents caused by SI products. Meetings are to be held regularly to monitor implementation by government departments.

Sources 1

This is just a teaser for now: the exact memory capacity and price have not been disclosed, and the full Lushan unveiling is set for December 29 at the MUSA Developer Conference 2026.

Moore Threads promises Lushan graphics card with record memory capacity for SI

This is just a teaser for now: the exact memory capacity and price have not been disclosed, and the full Lushan unveiling is set for December 29 at the MUSA Developer Conference 2026.

Chinese company Moore Threads has published a teaser for a new consumer graphics card codenamed Lushan. The company promises a record video memory capacity for the consumer segment — more than the 32 GB of the GeForce RTX 5090. It also claims up to a 64-fold performance increase in SI tasks compared with previous solutions. The card is positioned for both gaming and running SI models locally, and will get a new UNITE rendering architecture and next-generation hardware ray tracing. The exact memory capacity and prices have not been disclosed. The full unveiling is set for December 29 at the MUSA Developer Conference 2026.

Sources 1

Robots in Australia are now served by a local cloud node and an on-site team, but AGIBOT did not disclose delivery volumes or customer names.

AGIBOT opens cloud node in Australia to run its robots

Robots in Australia are now served by a local cloud node and an on-site team, but AGIBOT did not disclose delivery volumes or customer names.

AGIBOT has deployed a cloud node in Australia: it supports the local robot fleet, while an on-site team assesses sites and provides technical support. The company announced this in a press release on November 20. The deployment was presented at the Physical AI on AWS: ANZ Roadshow 2026 in Sydney on November 10, where AGIBOT general manager for Australia and New Zealand Gary Sun spoke. The company's robots were shown at the New Zealand–China Products Expo in Auckland on November 20–22. Commercial cleaning, security patrolling and premium services were named as priorities for New Zealand. The company did not disclose delivery volumes or customer names.

Developers can add speech synthesis in ten languages with intonation set by a text instruction, without their own infrastructure, at $20 per 1 million characters.

Qwen3 TTS lands on Together AI: instruction-based tone control and 97 ms to first audio

Developers can add speech synthesis in ten languages with intonation set by a text instruction, without their own infrastructure, at $20 per 1 million characters.

Together AI announced that the open speech synthesis model Qwen3 TTS is available to developers on Together Serverless Inference and on dedicated endpoints. The model separates voice from manner of speech: there are nine preset voices, each with its own native language, and intonation is set by a text instruction. Each voice speaks ten languages — Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian. According to Together AI, the inference stack is tuned so that the first audio segment arrives in about 97 milliseconds. HTTP streaming and WebSocket are supported, so text can be fed in as it is generated. The price is $20 per 1 million characters. The Qwen3 TTS 12Hz 1.7B weights were published by Qwen under the Apache 2.0 license.

Sources 1

The cut to the final remains very narrow — 60 people out of 69,000 applications — while winners and runners-up receive 100 Unified State Exam points in computer science.

All-Russian SI olympiad draws almost 70,000 applications, with 11 medals for schoolchildren

The cut to the final remains very narrow — 60 people out of 69,000 applications — while winners and runners-up receive 100 Unified State Exam points in computer science.

The awards ceremony for the winners and runners-up of the VI All-Russian SI Olympiad was held at the V. Tereshkova National Space Center. Russian schoolchildren received 11 medals. According to Deputy Prime Minister Dmitry Chernyshenko, a record 69,000 applications were submitted, while only 60 people reached the final. Finalists came from 18 regions of Russia, Kazakhstan, Armenia, Belarus and Cyprus; college and technical school students competed for the first time. The season's theme was "SI for the space sector", and the number of participating countries grew from 26 to 63. The absolute winner was 11th-grader Kirill Ivakhin from Penza Oblast. Winners and runners-up receive 100 Unified State Exam points in computer science; the competition is part of the SI federal project under the Data Economy national project.