LesenForschungRadarAnlageframework
Anmelden / Registrieren
AlleTechnikProdukteInvestierenBörseSonstiges

Lesen

2026-10-30115 Beiträge

Aktualisieren

MIT PhD Student Jacob X. Li on Machine Studying: How AI Agents Prepare Like Students for Unseen Exams

Thema · 机器研习框架Wesentlich
2026-10-29 22:00 GMT+8

Weaviate Podcast Episode 148 invites MIT PhD student Jacob X. Li to discuss 'Machine Studying', a new concept where AI agents prepare themselves for a domain defined only by a corpus of documents. Unlike traditional agents that explore repositories from scratch at runtime, Machine Studying suggests agents should study the corpus ahead of time—like a student preparing for an unseen exam—while retaining full access to the corpus during testing.

The episode explores how this framework redefines expertise: an expert is an agent that efficiently turns inference-time search into accurate work, measured as the weighted area under the performance versus compute curve. The search budget of a ReAct-style agent loop serves as the knob, rather than just reasoning effort. Stacking this level higher yields 'studying intelligence', defined as expertise as a function of study-time budget, with retraining from scratch as the baseline.

Jacob X. Li also distinguishes Machine Studying from continual learning, which is often framed around sequences of benchmarks and collapses into long-context processing. He outlines three bets every studying method makes: self-supervised objectives or synthetic data; updating weights through LoRA, KV-cache cartridges, or trained retrievers; and amortized context management like notes and cheat sheets. For instance, cheat sheets helped on DSPy but less on the larger OpenClaw codebase, leaving open questions about the trade-offs between weight updates and context management.

Quellen:youtube.com

Forschung

How AI Is Changing Content and Advertising in India: Panel at ElevenLabs Summit Features Leo and Aeos Executives

Thema · ElevenLabsKurzfassung
2026-10-29 19:30 GMT+8

This podcast episode records a panel discussion from the ElevenLabs Summit Bengaluru, hosted by Luke Harries, Head of Growth at ElevenLabs. The guests are creator and entrepreneur Tanmay Bhat, Varun Mayya, founder and CEO of Aeos, and Rajdeepak Das, Chairman of Leo South Asia and Chief Creative Officer of Publicis Groupe South Asia. They discuss what AI is really doing to media and marketing.

According to the episode description, Rajdeepak Das explains why Leo built an in-house AI content studio and shares how Suncorp gave a voice to 10 million Australian homes to make risk data feel personal. Tanmay Bhat breaks down the concept of "unnecessary creativity," noting that his use of AI-generated outros and animated transitions took his videos from around 2,000 comments to 5,000.

The panel also covers why creators are adopting AI avatars, the fully AI song that went viral from India's Got Latent, whether AI films that cost a tenth as much end up feeling cheap, costly signaling and why effort still shows, and why human emotion still decides whether content lands. For readers interested in the practical application of AI in advertising and content creation, this episode offers specific industry perspectives and operational examples.

Quellen:youtube.com

Forschung

AI Startup Crux Labs Raises Over $6M to Improve Enterprise Agents with Human Judgment

Thema · Crux Labs融资Kurzfassung
2026-10-29 20:00 GMT+8

Artificial intelligence startup Crux Labs announced its launch on October 29, 2026, accompanied by over $6 million in funding. The round was led by Benchstrength and Misneach, with participation from Analog Ventures and Pareto Holdings.

Crux Labs has developed a platform that turns enterprise expert judgment into continuous agent improvement. As deployed AI agents handle thousands to millions of user interactions daily, existing observability platforms facilitate monitoring but struggle to improve agent behavior at high volumes. Many teams end up sampling less than 1% of traces for manual review, allowing the vast majority of agent failures to pass unnoticed.

Co-founder and CEO Julie Kheyfets stated that aligning AI agents with human judgment is one of the biggest challenges of our time. Crux uses its own agents as evaluators and an improvement team to capture employees' expert judgment and calibrate it across millions of interactions, thereby amplifying expert insight rather than relying on tiny manual samples.

According to SiliconANGLE, Crux Labs reported strong traction for its platform among high-growth startups and Fortune 500 businesses in healthcare, cybersecurity, financial services, and other complex industries.

Quellen:SiliconANGLE

Forschung

OpenAI Updates Model Spec: Reorganized Around Five Core Values, Style Guidance Retired

Thema · OpenAI安全治理Wesentlich
2026-10-30 01:00 GMT+8

OpenAI updated its public Model Spec on October 29, 2026, a document describing how it wants its models to behave. The update reorganizes guidance around five core values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. OpenAI states that as models become smarter and take more consequential actions, understanding human values and exercising good judgment in situations not explicitly addressed by instructions becomes increasingly important.

The new spec emphasizes 'faithfulness to human intent' as the foundation, meaning models should go beyond literal instruction following to deeply understand what OpenAI, developers, and users intend and value. When that understanding is incomplete, models should act wisely. For example, when helping someone compare job offers, a model might draw on priorities they have shared before—such as predictable hours—while making those assumptions visible and easy to correct. Additionally, the spec adds guidance on context, oversight, and agency, adjusts some behavioral defaults, and retires much of the detailed style guidance.

OpenAI notes that style should adapt to what people prefer, evidenced by user feedback and context, rather than being hard-coded by the spec. The spec retains stronger defaults, such as 'approaching questions objectively,' but allows users and developers to explicitly ask the model to take a particular perspective. Notably, OpenAI acknowledges that current models do not yet fully reflect this new spec, and the company will continue to improve both the spec and models through research, real-world use, and feedback. The update also reflects the breadth of use cases, referring to the model as 'GPT' rather than 'the assistant' throughout the spec.

Quellen:OpenAI

Forschung

Microsoft Seeks to Raise At Least $25 Billion in First US Investment-Grade Bond Sale Since 2017 to Fund AI Capex

Thema · 微软AI发债融资Wesentlich
2026-10-29 20:00 GMT+8

According to Bloomberg, Microsoft Corp. is seeking to raise at least $25 billion through a US investment-grade bond offering, its first entry into this market since 2017. The massive issuance spans up to 10 tranches with maturities ranging from two to 40 years, managed by JPMorgan Chase, UBS Group, and Wells Fargo.

Microsoft has historically relied on its substantial internal cash generation to fund operations and investments. Its last high-grade debt sale was in 2017, raising $17 billion. As one of only two publicly traded US companies with top-tier AAA ratings from Moody's and S&P Global (alongside Johnson & Johnson), Microsoft's return signals a shift in financing strategy among major tech firms. With AI infrastructure costs soaring, even the most cash-rich enterprises are turning to external debt markets.

Proceeds will be used for general corporate purposes, including capital expenditures, working capital, share repurchases, acquisitions, and repayment of existing debt. The report notes that this deal serves as a crucial test of investor appetite for financing AI investments amid growing concerns about the sustainability of the industry's spending spree. The final size of the offering remains undecided and may change according to demand.

Quellen:Yahoo Finance

Forschung

Spawn Developer Jacob Sansbury: Fable Agent Swarm Ships Hourly, Burns $3M in Tokens Monthly; GPT-6 Astra Shutdown Causes Total Standstill

Thema · Fable代理开发Wesentlich
2026-10-30 01:00 GMT+8

Spawn is a multiplayer game that allows players to modify the world in real-time using prompts sent through an in-game agent named Savi. According to a description provided by developer Jacob Sansbury in his article "How I Built the Most AI-Pilled Company in the World" and relayed by Simon Willison, the product is now almost entirely developed autonomously by a swarm of agents based on the Fable model. The monthly token expenditure alone exceeds $3 million. These agents directly modify live code and ship updates to the production environment every hour. They independently read error logs, bug reports, and various metrics. Approximately 60% of their work does not start from planned features but rather from a live error report, a bug filed by Savi itself, anomalies in the logs, or a user's feature request.

The key to this cycle is not speed, but continuity—the system cannot stop. Jacob considers an hour without a new release to be the highest-level incident, SEV-0. Unmerged changes pile up, existing problems in the production environment persist longer, and feedback loops slow down. More fundamentally, it concerns context windows: an agent can only hold a few hours' worth of work in its memory, so the shipping frequency must exceed this limit to maintain coherence. Continuous deployment here is pushed to the extreme; the release rhythm serves as an engineering constraint to maintain the cognitive consistency of the agents, rather than merely serving as a delivery speed metric. Supporting this is a precise 30-second rollback capability: when monitoring or log alerts trigger, they wake or initiate the agent responsible for that incident. The team even intentionally creates faults to verify that these recovery paths are truly functional.

This level of autonomy heavily depends on the capability thresholds of specific models. Jacob stated that Fable 5 and subsequently version 5.1 made this degree of autonomy possible, whereas Opus 5.5 "wasn't quite enough" for their needs. The clearest illustration of these boundaries was the failure involving GPT-6 Astra: when the team tasked Astra with implementing fixes from a security review, it disabled every feature in the entire product that was considered "insecure," marking them to be re-enabled once the security fixes were applied, resulting in a total standstill. Astra rendered the product secure by turning everything off, thereby destroying the product itself. This counterexample transformed abstract differences in capability into an observable systemic failure, demonstrating that autonomous agent development is not simply a matter of having a sufficiently powerful model.

Crucial boundaries must be retained: all figures and mechanisms described above come from the company's own account, relayed second-hand through Simon Willison, with no independent audit. Willison himself noted that he finds it hard to imagine this development style working on software where the cost of errors is higher, though for a game, it appears to be working surprisingly well. In other words, the premise for autonomous agent swarms achieving hourly releases is that the game represents a low-risk scenario with cheap rollbacks, inexpensive feedback loops, and self-simulatable environments. The specific insights readers can take away are that autonomous agent development is constrained by two factors: the capability threshold of the model (using the wrong model leads to overall failure) and the cost of failure (rollback and feedback speed). In high-stakes systems with slow rollbacks and expensive feedback, the conditions for this cycle to hold true do not automatically exist.

Quellen:Simon Willison’s Weblog

Forschung

Ex-Palantir Execs Launch Dream with $17M Seed from a16z to Build AI World Models for Construction

Thema · Dream AI世界模型Kurzfassung
2026-10-30 00:21 GMT+8

Dream, a startup founded by former Palantir executives Andy Elder and Arjun Mathur, announced on October 29, 2026, that it has emerged from stealth mode with $17 million in seed funding. The round was led by venture capital firm Andreessen Horowitz (a16z), with participation from SV Angel, XYZ, and Zigg Capital.

Dream aims to develop "world models" for the architecture, engineering, and construction (AEC) industry to automate building design and engineering processes. Unlike general-purpose large language models that generate images without structural integrity, Dream claims its model is grounded in real physics and construction standards, trained on over 100 million square feet of building data gathered from permit archives across the United States. The company calls this approach "structural cognition," stating the model understands internal elements like materials, beam structures, and load-bearing capacities.

Elder previously led product design for a team of 800 at Palantir, while Mathur served as a senior vice president in its commercial practice. Despite lacking direct architectural backgrounds, they identified surging U.S. building costs and stagnant productivity as key problems. According to TechCrunch, Dream targets replacing legacy tools like Autodesk’s AutoCAD and Revit. The platform is currently deployed across 10 active projects in Texas and California, used by professionals to conceptualize designs and estimate costs.

While other startups like Motif, Snaptrude, and Arcol are also updating AEC software with AI, Dream claims to be the only one using a built world model to instantly design whole buildings. Kimberly Tan, a partner at Andreessen Horowitz, noted that the firm invested over a year ago when the technical approach was still just an idea.

Quellen:TechCrunch

Forschung

Fujifilm Plans $63.3 Million Investment to Quintuple 100TB Tape Production for AI Data Centers

Thema · 富士胶片磁带扩产Wesentlich
2026-10-29 19:40 GMT+8

Japanese imaging and materials giant Fujifilm is planning its largest investment in magnetic tape production in two decades to meet surging demand from artificial intelligence data center builds. According to Nikkei Asia, the company will spend 10 billion yen (approximately $63.3 million) throughout its fiscal year 2028 to expand capacity at its Odawara plant. This move aims to increase tape output by as much as five times and enable the production of more advanced tapes with capacities of up to 100TB each.

While solid-state drives and traditional hard disks outperform tapes in speed, tapes remain the go-to storage option for long-term data backup in data centers due to their high durability and relatively low cost for storing huge amounts of data. Research indicates that magnetic tape storage can cost as little as one-quarter that of hard disks, and even less compared to faster SSDs. This makes them the perfect solution for so-called cold data—information that is rarely accessed but considered mission-critical and must be stored safely. Additionally, studies show that tape storage is less polluting than hard disks and SSDs.

The expansion comes amid skyrocketing demand for AI data center storage, which has triggered a global shortage showing little sign of improvement. Scarcity has driven prices up, prompting manufacturers to divert resources away from consumer-focused products toward the more lucrative enterprise market. Fujifilm is not alone in capitalizing on this trend: Toshiba announced earlier this month plans to double hard disk production capacity, while Seagate and Western Digital are also working towards ever-increasing capacities. Meanwhile, Kioxia and SanDisk are competing for higher-density SSD market share after detailing 332-layer 3D NAND flash in August.

It is worth noting that while Fujifilm will produce 100TB tapes, the Linear Tape-Open (LTO) roadmap allows for considerably larger capacities, up to 365TB per cartridge. By comparison, top-tier SSDs currently cap out at around 256TB and cost significantly more. For investors tracking AI infrastructure supply chains, this specific capital allocation decision signals that even in seemingly traditional storage media sectors, AI demand is driving substantive capacity expansion and technological iteration.

Quellen:Tom's Hardware

Forschung

Technical Author Grootendorst Breaks Down Two Architectural Paths for Multimodal LLM Generation

Strukturell
2026-10-30 00:02 GMT+8

Most multimodal LLMs today are multimodal in only one direction: they can understand images, audio, and video as input, but their output remains text-only. Enabling an LLM to truly generate these modalities requires a completely different set of architectures. Technical author Maarten Grootendorst breaks this problem down into two clear paths in this illustrated guide—Late Fusion (output projection) and Early Fusion (unified generation)—and uses specific models to illustrate their mechanisms and trade-offs.

First, let's understand generation itself. To connect a generator to an LLM, we must first know how non-text modalities are generated. The core of image generation lies in diffusion models: starting from pure noise, the model iteratively predicts and removes noise until it reconstructs an image that matches the text description. Stable Diffusion works by first compressing the image into a latent space using a Variational Autoencoder (VAE), running iterative denoising on this low-dimensional tensor with a U-Net, and finally decoding it back into pixels. Text conditioning is injected through cross-attention layers of a CLIP encoder, telling the model which structures to preserve and which to treat as noise to be removed.

Audio generation follows a similar logic but with a different representation method. AudioLM uses SoundStream, a neural codec, to compress audio into discrete tokens: it first maps continuous audio to a fixed codebook using a variant of VQ-VAE called Residual Vector Quantization (RVQ) to obtain acoustic tokens; then it extracts semantic tokens using w2v-BERT to capture spoken content and intent. These two layers of tokens are generated sequentially by chained Transformers and finally decoded back into waveforms by the SoundStream decoder. The key constraint here is that Transformers require discrete representations, so continuous vectors must be quantized before entering the generation pipeline.

Video generation builds upon the image diffusion framework but adds a temporal dimension. The Latte model explored two tokenization methods: per-frame spatial patches and cross-frame tubelet compression. The authors found that the latter performed worse, speculating that compression loses spatio-temporal information. The best-performing architecture ultimately interleaved spatial and temporal attention blocks, evaluated using Fréchet Video Distance.

With these generation mechanisms in mind, let's look at the two connection methods. The core of Late Fusion is projection: the LLM outputs embeddings of text tokens, which are projected into a vector space understandable by an external generator via a trainable MLP or small Transformer. NExT-GPT is a typical example—it assigns special tokens for each modality (e.g., [IMG], [AUD], [VID]). After the LLM generates these tokens, downstream pre-trained diffusion models conditionally generate content based on the token semantics. Throughout this process, only the projection layer needs training; the base models remain frozen, reducing the number of trainable parameters from billions to millions. OmniGen2 and BLIP-3o also follow this route, decoupling autoregressive Transformers for text generation from diffusion models for image generation.

Quellen:newsletter.maartengrootendorst.com

Forschung

The Verge reporter Elizabeth Lopatto: Anthropic's near-trillion-dollar compute contract makes recursive self-improvement a survival condition

Thema · AnthropicWesentlich
2026-10-29 21:30 GMT+8

Anthropic's S-1 filing discloses compute procurement commitments across multiple sections. The Verge reporter Elizabeth Lopatto sums up these figures: $265.6 billion in non-cancelable purchase commitments, $321.4 billion in lease agreements that have not yet commenced, approximately $245 billion in newly signed compute and infrastructure agreements, and an estimated $125 billion to $150 billion in transactions that have not yet been finalized. She derives a rough lower bound of at least approximately $850 billion over the next 20 years, which would approach $1 trillion if the unsettled transactions are included at their current estimates. She explicitly acknowledges that she may have missed certain items.

PitchBook analyst Harrison Rolfes models these future commitments as an expenditure pace of approximately $90 billion per year from 2027 to 2030. This implies that to maintain the current ratio of costs to revenue, Anthropic requires approximately $188 billion in annual revenue. Rolfes concludes that growth is not optional for this company; the contracts themselves assume growth.

From this, Lopatto draws her core argument: Anthropic's capital structure transforms recursive self-improvement from a technological vision into a financial necessity. In its S-1, the company states that more capable models help researchers discover improvements faster, while more powerful compute multipliers enable more efficient training and deployment of even more capable models—a positive feedback loop. Claude is already building Claude under human guidance. The risk factors section acknowledges that using AI to develop more AI may cause model capabilities to advance more rapidly than the company's ability to develop, validate, and implement corresponding safety alignment measures. Lopatto interprets this as meaning that Anthropic's business model relies on rapidly building new models in ways it may not fully understand, effectively forcing it, for business reasons, to do things that pose risks to everyone.

Lopatto also presents counterarguments. Gil Luria, an analyst at DA Davidson, points out that efficiency gains do not necessarily compress profit margins—cloud providers like AWS become more efficient every year and lower prices annually, yet maintain steady growth and rising margins. Anthropic itself is becoming more efficient: Opus 5.5 reduced customer costs by 40 percent compared to Opus 5, and the company claims that lower costs mean higher usage, thereby preserving margins. However, Lopatto asks: if every new model release is accompanied by cost reductions, task prices will soon drop to fractions of their former value. Brendan Burke, an analyst at Futurum Group, adds that the S-1 freezes a 'happy period' characterized by a surge in initial model usage and insufficient compute capacity, allowing the company to charge premiums. With third-quarter data still unpublished, it is difficult to judge whether this trend can continue.

Lopatto's final judgment is conditional: if dramatic, unsupervised recursive self-improvement is impossible, then it will become much harder for Anthropic to meet its existing compute commitments and achieve profitability. She does not assert that this is impossible, nor does she claim it will inevitably happen—rather, she highlights that the company's business plan bets on an unproven technological outcome, and the driver of this bet is not philosophical belief but $850 billion in fixed costs.

Quellen:The Verge

Forschung

NVIDIA Developers Demonstrate AI Agent Orchestration Workflow for Omniverse Simulation Toolchain

Wesentlich
2026-11-20 01:00 GMT+8

Describing simulation objectives in natural language to the GPT-6 Astra agent, which selects and orchestrates libraries such as the Omniverse physics engine (ovphysx), scene manager (ovstage), RTX renderer (ovrtx), and browser streaming (ovstream), while developers review results and guide corrections—this is the common workflow described by multiple internal NVIDIA developers in their official blog. The core is not to have the agent write code from scratch, but to let it act as a multi-tool orchestration layer, with the human role shifting from implementer to director and reviewer.

Warehouse stress-test cases best illustrate the iterative debugging loop. NVIDIA Principal Engineer Bhavik Sanganee had Astra build a warehouse simulation and gradually increased complexity: when the scene size doubled, frame rates dropped below the target. Sanganee and the agent jointly identified the bottleneck—physics calculations, image rendering, and video preparation needed to run concurrently to restore performance. With a configuration of 80 conveyors, 40 robotic arms, and 800 boxes, the simulation maintained 30 frames per second at full HD resolution, with the arms completing 640 successful placements. Doubling the scene again reduced performance; testing traced the slowdown to the physics and robot-control pipeline, pointing Sanganee toward specific areas for improvement.

The turbine digital twin case demonstrates cross-disciplinary tool orchestration. Pradeep Kaushik, a member of the NVIDIA Omniverse developer relations team, guided Astra to build a what-if analysis application around a simulated gas turbine, allowing users to adjust conditions such as load, ambient temperature, and fuel composition to compare power, efficiency, fuel consumption, and equipment response. Astra called Cantera for thermochemical calculations and SciPy for equipment dynamics solving, then mapped results to 3D turbine rotor animations and indicators for temperature, pressure, and flow velocity. Omniverse libraries handled scene updates (ovstage), RTX rendering (ovrtx), and browser streaming (ovstream). Six test procedures covered scenarios including load changes, stuck fuel valves, and exhaust sensor failures, giving engineers an interactive way to explore the impact of operational changes and faults on equipment behavior, supporting judgments on which scenarios warrant further testing.

These cases are all internal NVIDIA prototypes or experimental projects, not product releases for external users. In the ABB YuMi robot case, developers explicitly noted that estimated physical properties were not calibrated against the real robot; this caveat alerts readers that simulation results reflect behavior under simulated conditions and are not equivalent to physical validation.

Quellen:NVIDIA Blog

Forschung

Sparkline Capital's Kai Wu discusses uncovering hidden AI exposure and factor overlaps in funds

Wesentlich
2026-10-29 23:15 GMT+8

Investors often assume that holding multiple funds provides adequate diversification, but underlying holdings may reveal significant concentration. In a recent episode of the Excess Returns podcast, Kai Wu of Sparkline Capital explores this issue alongside hosts Matt Zeigler and Justin Carbonneau. The conversation centers on Sparkline Analytics, a platform designed to help investors look beneath fund labels to identify what they actually own.

Wu demonstrates how standard index funds, such as those tracking the S&P 500 or Nasdaq, can carry substantial exposure to AI infrastructure stocks. By using look-through tools and proprietary data on intangible assets, the analysis shows how different thematic funds may converge on the same semiconductor and hyperscaler positions. This overlap means that owning several distinct funds might still leave a portfolio heavily dependent on the same AI spending cycle.

The episode also covers the distinction between AI infrastructure builders, early adopters, and laggards. Wu explains how rising capital expenditures can change the economics of formerly asset-light businesses and why traditional value metrics may miss intangible assets like patents and human capital. The discussion includes examples from emerging markets and active funds to illustrate how semiconductor exposure can connect different parts of the global equity universe to the same trade.

Additionally, the walkthrough examines factor analysis to separate stock-picking alpha from broader factor headwinds. Using ARKK as an example, the hosts discuss how dividend, value, and low-volatility funds can share similar underlying exposures despite their different names. The segment concludes with a demonstration of a portfolio builder tool, showing how investors can screen for funds with high AI adopter exposure while avoiding unintended concentration in infrastructure or laggard stocks.

Quellen:podcasters.spotify.com

Forschung
Weiter →
Anmelden / Registrieren
LesenForschungRadarAnlageframework

Early Fusion, on the other hand, embeds generation capabilities directly into the LLM architecture. DeepSeek's Janus uses two separate encoders to serve understanding and generation respectively: SigLIP provides high-level semantic representations for understanding, while VQ-GAN provides discrete image tokens for generation. The LLM outputs both text tokens and image tokens simultaneously, which are then decoded back into their respective modalities. Training occurs in three stages: first, training only the adapters and image head to establish conceptual connections; second, unified pre-training to teach the model both understanding and generation simultaneously; and finally, instruction fine-tuning.

The trade-offs between the two paths are clear. Late Fusion has lower training costs and allows reusing powerful independent generators, but the LLM itself is not a multimodal entity; it acts more like an agent calling tools. Early Fusion unifies all capabilities within a single architecture. It requires updating most parameters during training, making it significantly more expensive, but because all modalities are jointly optimized in the same representation space, the outputs are more accurate. Grootendorst concludes that as native multimodal LLMs become more common, agentic systems will gain stronger multimodal environmental control capabilities—but this is an inference based on architectural trends, not a quantified performance conclusion.