LesenForschungRadarAnlageframework
Anmelden / Registrieren
AlleTechnikProdukteInvestierenBörseSonstiges

Lesen

2026-11-2154 Beiträge

Aktualisieren

METR report: AI models excel at coding but lag in judgment tasks

Wesentlich
2026-11-21 02:15 GMT+8

The AI evaluation nonprofit METR has released an update to its widely discussed "time horizon" plot. The report finds that while frontier AI models continue to improve rapidly on tasks with clear success criteria—such as software engineering and mathematics—they struggle significantly with open-ended tasks that require human-like judgment, such as designing research evaluations or scoring philosophical arguments.

METR's original time horizon plot measured task difficulty by the time it took humans to complete them, showing that the length of tasks AI could successfully complete was doubling every four months. The updated report expands this suite to include more complex, long-duration, and judgment-heavy tasks. On these new metrics, even the most advanced models failed to produce evaluations deemed useful by METR researchers, and their performance on philosophical argument scoring remained below human levels.

This divergence highlights a key bottleneck in automating AI research and development (AI R&D). Progress on "verifiable" tasks is driven by reinforcement learning, where systems can be trained through repeated attempts and automated feedback. However, applying this technique to tasks where success is ambiguous or takes longer to ascertain remains difficult. Experts note that while models are getting better at these fuzzy tasks—with some recent models achieving "borderline above the bar" performance compared to "clearly below" a year ago—the lack of robust metrics makes predicting future progress challenging.

The report also addresses practical difficulties in evaluation. For many of the new tasks, human completion times are estimated rather than measured due to cost and expertise constraints. Additionally, subjective tasks often rely on limited human raters, raising questions about the generalizability of the scores. Despite these limitations, the inclusion of judgment-based tasks is seen as a necessary step toward understanding whether AI can autonomously drive scientific discovery.

Quellen:MIT Technology Review

Forschung

Anthropic IPO: Shares Open at $111.50, Close at $97, Drop to ~$85 Next Day

Wesentlich
2026-11-21 02:17 GMT+8

AI company Anthropic completed its initial public offering (IPO) on the Nasdaq exchange in New York, United States. The event was marked by a historic debut that failed to sustain momentum, reflecting mixed investor sentiment towards the broader artificial intelligence sector.

Shares began trading at approximately 18:00 local time on Thursday evening at $111.50 per share. This opening price represented a 17% premium over the fixed offering price of $83. With around 980 million shares issued, the total proceeds were considered record-breaking for an AI-focused listing. During the first day, the market capitalization briefly surpassed the $2 trillion threshold.

However, the stock price declined rapidly after the initial surge. It fluctuated in the $90 range for the remainder of the first trading day, closing at $97 with an intraday low of $92. On the following day, the share price dropped further, falling close to the original offering price of $83 before stabilizing at around $85.

The IPO demand was exceptionally high prior to listing. Reports indicate the shares were oversubscribed at least six times, meaning demand was sixfold the available supply. Institutional investor demand reportedly reached approximately $400 billion, while retail investor interest was cited as nearly $100 billion, according to anonymous sources referenced by CNBC.

Analysts remain divided on the long-term implications. While the debut was superlative in terms of volume and initial pricing, it did not replicate the sustained rocket-like rise seen in other recent tech listings such as SpaceX. The decline on the second trading day is attributed by some to typical profit-taking effects common in hot IPOs, though others view it as a signal of cooling enthusiasm.

The broader market context contributed to the uncertainty. Global markets showed signs of unrest, with Taiwan's stock exchange heading toward its worst week since 1991 and semiconductor giant AMD experiencing significant losses. These factors suggest that Anthropic's listing did not fully dispel nervousness within the AI sector.

Key upcoming events will influence future performance. Insider selling restrictions are expected to lift in approximately 180 days, which could increase volatility. Additionally, the next quarterly earnings report is scheduled for February. Until then, most analysts are refraining from making definitive forecasts about the stock's trajectory.

Quellen:heise online

Forschung

Meta Raises Digital Brain Project Data Target to 25,000 Hours

Kurzfassung
2026-11-21 02:30 GMT+8

Meta announced on November 20, 2026, that it is raising the data collection target for the Digital Brain Project from 15,000 hours to 25,000 hours of de-identified brain data. Launched with the Rothschild Hospital Foundation and the University of Montreal, the project funds researchers to collect and openly share neural data to help understand human intelligence and improve neurological health.

The announcement was part of the highlights from Meta's Open Source AI for Science Summit held at its Menlo Park headquarters. While the summit also featured other tools like UMA Nitro for chemistry simulations and ATLAS for mathematical proofs, the Digital Brain Project represents a concrete step in building shared infrastructure for AI-driven medical research. By increasing the volume of openly available neural data, Meta seeks to lower barriers for scientists using AI to tackle complex problems in neurology.

Quellen:Meta Newsroom

Forschung

Google Gemini cuts wildfire radio transcription error rate to 22.34%, enabling sub-30-second alerts

Wesentlich
2026-11-21 01:00 GMT+8

Google has announced a new AI-powered tool built with its Gemini model that helps the nonprofit wildfire alert service Watch Duty transcribe noisy emergency radio communications faster and more accurately. The collaboration, supported by a Google.org Fellowship, aims to reduce the time it takes for volunteers to identify new fire incidents from decades-old analog VHF radio feeds.

The core technical achievement is a significant reduction in Word Error Rate (WER) on degraded audio. According to Google, the system reduced WER from 38% down to 22.34% when processing static-heavy VHF radio audio containing specialized firefighting terminology. This improvement allows the tool to flag potential new dispatches in less than 30 seconds, compared to the manual monitoring of up to 20 feeds previously required by reporters.

The practical impact was demonstrated during the Railroad Fire in Washington state in July 2026. When a 911 dispatch went out over local fire radio, the Gemini-powered tool transcribed it directly from the noisy audio and flagged it within 30 seconds. This head start allowed Watch Duty’s team to verify the incident and push a Critical Alert—which bypasses do-not-disturb modes—before flames reached homes. It marked the first time Watch Duty had sent a nighttime Critical Alert.

The system currently transcribes more than 200 hours of radio speech daily from approximately 400 active feeds and is designed to handle over 10,000 feeds simultaneously. Google reported that volunteer reporter satisfaction with radio transcription tools increased from 35.3% to 92.3% as a result of the collaboration. The tool was trained on more than 20,000 hours of past radio recordings.

Quellen:Google

Forschung

AMD Releases Nitro-P: Pixel-Space Text-to-Image Model Achieves 90 Samples/s in One-Step Generation

Wesentlich
2026-11-20 08:00 GMT+8

On November 20, 2026, AMD released Nitro-P, an end-to-end pixel-space text-to-image (T2I) diffusion framework. Built upon its previously released latent-space model Nitro-E, Nitro-P extends the family to a VAE-free design by adopting a hierarchical semantic-to-pixel architecture. This structure consists of a patch-level semantic modeling component based on E-MMDiT and a pixel-level modulation-guided reconstruction component implemented as a lightweight U-Net decoder.

The core advantage of operating directly in pixel space is the removal of the Variational Autoencoder (VAE) decoder from the inference pipeline. In traditional latent-space models, generated latent representations must be converted back to RGB images by a VAE decoder, which incurs non-negligible computational overhead. Nitro-P bypasses this final decoding pass entirely. According to AMD's self-reported benchmarks on a single AMD Instinct™ MI300X GPU with a batch size of 32, the full-step Nitro-P model achieves a throughput of 10.26 samples/s for 512px image generation. Its distilled one-step variant further boosts throughput to 90.02 samples/s.

To isolate the effect of removing the VAE decoder, AMD compared the end-to-end throughput of Nitro-P against Nitro-E. The results show that the performance gap is largest in the few-step regime. At 512×512 resolution, the one-step distilled Nitro-P reaches 90.02 samples/s, versus 44.13 samples/s for Nitro-E. However, as the number of sampling steps increases, the gap narrows. At the full 20-step setting, Nitro-E is actually faster (18.83 samples/s vs. 10.26 samples/s), because the denoising backbone dominates runtime and the latent-space path remains cheaper per step. These findings indicate that Nitro-P is particularly well suited to low-NFE (Number of Function Evaluations), high-throughput serving scenarios.

In terms of generation quality, Nitro-P achieves competitive GenEval scores of 0.88 at 512×512 and 0.87 at 1024×1024. The one-step distilled variant sees a moderate drop in quality, scoring 0.85 at 512×512. All models were trained from scratch on public data, including a curated dataset of 25 million real and synthetic images. AMD has released both the model weights and the complete training code to support reproducibility.

Quellen:ROCm Blogs

Forschung

Anthropic mathematician Levent Alpöge claims to have discovered a proof of the Riemann Hypothesis using Claude

Wesentlich
2026-11-21 00:30 GMT+8

Levent Alpöge, a mathematician employed by the US-based AI company Anthropic, has written in a paper that he believes he has discovered a proof of the Riemann Hypothesis while working with the company's AI model, Claude. The claim was reported by Tom's Hardware on November 20, 2026. The Riemann Hypothesis, proposed in 1859, is one of the most famous unsolved problems in mathematics and concerns the distribution of prime numbers.

Alpöge stated, "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis." He also suggested that the method could extend to the Generalized and Extended Riemann Hypotheses. However, this is a first-party claim from an individual researcher and has not yet been independently verified or peer-reviewed by the broader mathematical community. Proving such a fundamental theorem typically requires extensive scrutiny over time.

This event highlights the growing role of AI in scientific discovery, following earlier reports of OpenAI assisting with solutions to other complex STEM problems. While the potential for AI to aid in major breakthroughs is significant, readers should treat this specific claim as an attributed statement pending independent confirmation.

Quellen:Tom's Hardware

Forschung

Software Company Bending Spoons Acquires Mailchimp from Intuit for $3.68 Billion

Wesentlich
2026-11-21 00:16 GMT+8

Software company Bending Spoons has acquired the email marketing platform Mailchimp from US tax preparation giant Intuit for $3.68 billion. The transaction marks a significant loss for Intuit, which originally purchased Mailchimp in 2021 for approximately $12 billion.

Intuit stated that the sale allows it to sharpen its focus on AI tools for small businesses, with the proceeds intended for share repurchases. This decision comes as Intuit's stock price has fallen by more than half over the past year, pressured by market concerns regarding the impact of AI on traditional SaaS business models.

The deal structure is primarily cash-based, though Intuit will purchase $368 million in newly issued Bending Spoons shares, resulting in a net split of 90% cash and 10% stock. For Bending Spoons, this acquisition continues its strategy of consolidating late-stage web properties such as AOL, Eventbrite, Airtable, and Miro, aiming to improve their operational efficiency through centralized engineering and revenue systems.

Quellen:TechCrunch

Forschung

AMD's Alleged Circular Financing Reveals a Structural Vulnerability Across AI Chip Companies

Wesentlich
2026-11-20 23:44 GMT+8

Short-seller Hunterbrook published a report titled "Bank of AMD" alleging that the AI chip company has run an extensive customer-financing-and-investing operation featuring "all of the circular financing of Nvidia, but without the margins or pricing power." A separate Hunterbrook story claimed AMD delivered updated chip firmware to multiple blacklisted Chinese entities. AMD said the reports didn't properly represent its business, but shares fell almost 10% on the news.

The circular financing thesis is not new. Chip companies invest in their customers, who then use those funds to buy chips from the same companies. Revenue looks strong, but a meaningful portion is self-reinforcing rather than driven by independent end-demand. The analogy to the telecom bubble of the early 2000s is that vendors extended credit to customers who used it to buy vendor equipment, inflating apparent demand until the cycle broke.

What makes the AMD case specifically interesting is the margin differential. Hunterbrook's core claim is that AMD replicates Nvidia's circular financing structure but lacks the same pricing power and margins. Nvidia commands premium pricing and has historically maintained wider gross margins, giving it more room to absorb the inefficiency of circular deals. AMD operates at thinner margins, meaning the same financing structure consumes a larger share of its economic value. If a customer's ability to pay depends on the vendor's own investment, the revenue quality is structurally weaker when margins are thinner.

The China dimension adds a second vulnerability. Like Nvidia, AMD sells chips to Asian neocloud operators that serve China's big tech companies. New US government export control rules put those revenues at risk. Hunterbrook's separate claim about firmware delivered to blacklisted Chinese entities, if substantiated, would compound regulatory exposure beyond what export controls alone create.

Broader market behavior supports the concern without proving it. Big Wall Street investment firms that promised in August to provide approximately $500 billion in loans for data centers have yet to dispense a single dollar, according to the Wall Street Journal. Nvidia's interim CEO Colette Kress signaled she intends to dial back Jensen Huang's aggressive dealmaking, which included a record 121 investments through Nvidia and NVentures in 2026. Anthropic's IPO went off better than expected, yet across the AI industry, a lot of people are selling on the news.

The strongest objection to the circular financing thesis is that genuine end-demand may be growing fast enough to outpace the self-reinforcing component. If AI compute demand is real and expanding independently of vendor financing, then circular deals are a convenience rather than a dependency. The telecom analogy breaks if the underlying use case generates sufficient standalone economics to sustain the cycle without vendor support.

The conditions under which this thesis matters most: when export controls tighten further, when margin compression makes circular deals unaffordable, when interim management at major chip companies reduces dealmaking pace, or when data center loan commitments remain unfunded for extended periods. Each of these would test whether revenue reflects genuine demand or capital recycling. None have yet triggered a full cycle break, but the directional signals are visible in the market's reaction to the Hunterbrook report itself.

Quellen:newcomer.co

Forschung

EU Commission and Eight Member States Launch EDU-EDIC to Build Sovereign Digital Education Infrastructure

Wesentlich
2026-11-20 18:30 GMT+8

On November 20, 2026, the European Commission and eight member states—Belgium, Croatia, Finland, France, Greece, Slovenia, Spain, and the Netherlands—launched the European Digital Infrastructure Consortium on Education (EDU-EDIC). The initiative aims to pool resources to develop shared, secure, and interoperable digital infrastructure and services for education across the European Union.

The consortium is designed to strengthen Europe's technological resilience while making high-quality digital tools accessible to schools, teachers, and learners. It seeks to narrow the digital divide by supporting solutions that can be used across national education systems regardless of their local resources or technological capacity. Roxana Mînzatu, Executive Vice-President for Social Rights and Skills, Quality Jobs and Preparedness, stated that a priority is ensuring education is ready for the AI era. She noted that the consortium will work with member states to pool expertise in strategic areas such as cloud infrastructure, AI for education, secure data exchange across educational institutions, and Europe's education technology (EdTech) sector.

Henna Virkkunen, Executive Vice-President for Tech Sovereignty, Security and Democracy, emphasized that the launch is an important step towards achieving Europe's technological sovereignty in a sensitive and strategic sector. She argued that Europe must rely on trusted, safe, and interoperable digital solutions that reflect European values and principles. The consortium represents a milestone in fostering innovation and excellence in education while contributing to the EU's broader technological sovereignty ambitions.

Quellen:Shaping Europe’s digital future

Forschung

Over 40 US Ballot Measures Passed to Limit AI Data Center Growth

Wesentlich
2026-11-21 00:43 GMT+8

Following the 2026 US midterm elections, more than 40 ballot measures seeking limits on data center growth have passed across the country. This outcome represents a significant shift in the regulatory landscape for artificial intelligence infrastructure, moving beyond federal legislative debates to direct state-level constraints.

The passage of these measures highlights a growing public backlash against the rapid expansion of AI data centers, which are often criticized for their massive land use and energy consumption. For instance, Senator Mark Kelly recently described a specific data center project in New Mexico as being the size of "a thousand football fields," a claim verified by investigative outlet Searchlight New Mexico. Such scale has fueled local opposition, leading to successful voter initiatives aimed at curbing further development.

This trend poses a tangible challenge to tech companies planning large-scale AI deployments. While Silicon Valley executives may view this sentiment as a marketing or education problem, the concrete legal restrictions enacted through ballot measures create immediate operational hurdles. State lawmakers and governors are increasingly empowered to enforce these limits, potentially slowing down the pace of infrastructure buildout required for advanced AI models.

Quellen:semafor.com

Forschung

Suno Launches Visuals, a Video Tool Built on MiniMax's Open-Source H3 Model

Wesentlich
2026-11-20

US-based AI music platform Suno has announced the launch of Visuals, a new creative tool that integrates video and image generation into its music creation workflow. According to official disclosures, Visuals is powered by a model named H3 Move, which is built upon the open-source video model H3 developed by Chinese AI company MiniMax and fine-tuned specifically for music production scenarios.

Jack Brody, Chief Product Officer at Suno, stated that his team optimized the open-source model using proprietary data to better suit music video creation. While MiniMax's H3 provides the underlying foundation, Suno has adapted it for the music context, integrating songwriting, visual generation, and lip-syncing into a single process. The resulting videos are designed to reflect song structure and lyrics, supporting vocal synchronization. Users can generate clips up to 15 seconds long by combining music with text prompts or uploaded reference images, and further edit these via a timeline interface.

The move addresses previous limitations in video generation speed and cost, which had restricted such features to enterprise users. Brody noted that recent advancements in video models have made them viable for general consumers. MiniMax opened the weights of its H3 model in August, allowing developers to deploy and customize it. H3 is capable of understanding multimodal contexts (text, image, video, audio) and generating videos with native stereo audio at up to 2K resolution.

Visuals operates on a separate credit system from Suno's music generation. To protect intellectual property, Suno has implemented text and visual filters to block unauthorized use of specific artists' likenesses or copyrighted content, and will embed watermarks in shared outputs. This expansion follows a trend of AI platforms diversifying beyond their core modalities, occurring roughly six months after OpenAI discontinued its short-video tool Sora.

Quellen:量子位

Forschung

Anthropic mathematician Levent Alpöge claims to have solved the Riemann Hypothesis with Claude

Wesentlich
2026-11-21 00:34 GMT+8

Levent Alpöge, a mathematician at Anthropic, has claimed on his personal website that he has discovered a proof of the Riemann Hypothesis with the help of the company's AI model, Claude. The Riemann Hypothesis is one of seven Millennium Prize Problems posed by the Clay Mathematics Institute, concerning the distribution of prime numbers and considered one of the most important unsolved problems in mathematics since its proposal in 1859.

Alpöge stated that the method extends to resolving the Generalized Riemann Hypothesis and expects it to cover the Extended Riemann Hypothesis as well. He noted that he and Claude formalized the arguments without assumptions. However, the solution has not been independently verified, nor has it been accepted by the Clay Mathematics Institute. Under the institute's rules, a proposed solution must be published and withstand at least two years of scrutiny while receiving general acceptance from the global mathematics community before it can be considered for the $1 million prize.

Anthropic has not commented on or involved itself in Alpöge's announcement. In August, Anthropic had stated it was working on the Riemann Hypothesis with Claude but did not expect the techniques used to lead to a proof. This claim follows recent AI breakthroughs in mathematics, including OpenAI's widely discussed but still unverified work on the Navier-Stokes problem and other Millennium Prize challenges, which have sparked debate among researchers about verification bottlenecks and credit attribution.

Quellen:The Verge

Forschung
Weiter →
Anmelden / Registrieren
LesenForschungRadarAnlageframework