ReadingResearchRadarInvestment framework
Sign in / Sign up
AllTechnologyProductsInvestingPublic marketsOther

Reading

2026-11-2082 posts

Refresh

Three Retired V100 GPUs Outpace $5,399 Mac Studio in Long-Context Inference for $1,042

Quick take
2026-11-19

According to Tom's Hardware, a homelab builder known as esologic benchmarked three used Nvidia Tesla V100 GPUs, finding that the ~$1,042 setup outperformed a $5,399 Apple M5 Max Mac Studio in inference speed for the Qwen 3.6 35B mixture-of-experts (MoE) model at long context lengths.

The tests, conducted using the llama.cpp framework, compared generation speeds across different quantization levels and context sizes. With 8-bit quantization (Q8), the three V100s achieved 100.8 tokens per second at 4K context, dropping to 72.7 tokens per second at 128K. The M5 Max Mac Studio, equipped with a 40-core GPU and 128GB of unified memory, was slightly faster at 4K (105.5 tokens per second) but fell behind at 64K and 128K, where the V100 cluster was approximately 19% faster at the longest context tested.

The V100s represent older 'e-waste' enterprise GPUs, whose advantage lies in combining VRAM capacity and bandwidth through multi-GPU parallelism, enabling larger models or longer contexts. While modern consumer devices like the Mac Studio may excel in short-context performance and energy efficiency, they were surpassed by the low-cost legacy hardware in this specific long-text workload. The builder also noted that the V100s consumed about half the energy of an equivalent number of P100s. Due to the use of different runtimes (llama.cpp vs. MLX), these figures should be taken as indicative rather than independently verified.

Sources:Tom's Hardware

Research

AWS Engineer Chambers: A Single Forward-Pass Decision Model Allows Agents to Skip Unnecessary Memory Retrieval

Quick take
2026-11-20 09:37 GMT+8

Say “good morning” to an AI agent, and it might search its long-term memory before saying hello back. Somewhere in that system, a service performs a semantic lookup to determine whether your greeting requires understanding your beverage preferences. Even if you take your coffee very seriously, this seems like a lot of infrastructure for two words.

Mike Chambers, a Senior AI DevEx Engineer at Amazon Web Services (AWS), presented a solution to The New Stack Editor-in-Chief Nick Lucchesi at AGNTCon + MCPCon North America 2026: using a decider model to handle the small judgments that accumulate inside an agent. These models are also known as decision models or System One models; they select from a fixed set of options or rate inputs, returning calibrated probabilities without generating free-form text. This concept was popularized by Jev of TypeSafe AI, and AWS’s Strands Labs released the open-source Strands Decider 2B, a model with two billion parameters, on October 1.

Chambers describes this approach as having an agentic-powered, language-aware if-statement within your code. Developers have used classifiers for years, but what distinguishes this generation of decider models is that general-purpose pretraining allows you to try them on new classification tasks without first training your own model. You provide natural-language context, ask a question, and receive a numerical score that the application uses to choose a branch. In the models Chambers describes, this score is returned in a single forward pass, with no token sequence requiring generation.

This makes decider models suitable for placement on the critical path of every request. Chambers experimented with the open-source Strands framework and Amazon Bedrock AgentCore Memory: he placed Strands Decider before the memory lookup to judge whether an incoming message required fetching information from past conversations. Greetings could pass straight through; requests dependent on what the user said earlier would still trigger a search. He also used a decider after retrieval: while a memory service can return the most relevant matches, some may be unrelated to the current task, so the decider can score each result for relevance before the agent incorporates it into the context.

One decision saves a service call, and another reduces the amount of irrelevant material the model must process. A model router can also use scores to send routine prompts to smaller models and reserve more capable models for complex tasks, saving resources from both the routing decision itself and the downstream work it avoids.

However, Chambers explicitly characterizes his memory experiment as experimental. While a number in a predictable JSON field makes it easy to integrate a decider into an application, determining whether that number is reliable enough to act upon requires testing.

"Will the number be correct? That’s for your testing to figure out," Chambers said. When stakes rise—such as when an agent is about to spend money rather than recommend a coffee—Chambers advocates for setting up checks outside the agent’s loop and requiring human approval for decisions that are difficult to reverse.

Sources:The New Stack

Research

Cognition Launches Devin Bots: AI Agents Autonomously Manage Engineering, Build Android App in 44 Days with Minimal Oversight

Material
2026-11-20 04:22 GMT+8

Cognition launched Devin Bots on Tuesday, introducing persistent agents designed to oversee multiple Devin coding sessions and take responsibility for engineering outcomes over days or weeks. CEO Scott Wu argues that coordination, rather than just coding, represents the next opportunity for automation.

The company disclosed a specific case study involving the development of Devin's Android version. Between September 28 and November 11, a span of 44 days, a single Devin Bot managed the project. During this period, the Bot merged 1,134 pull requests, wrote 344,651 lines of Kotlin code, and closed 1,217 tickets. The human engineer overseeing the project sent only 183 messages, with zero interactions recorded on 22 of those days.

The Android app is currently in open testing on Google Play, though Cognition notes it still requires work to match the iOS version. The company also highlighted other examples, such as a Layout Bot reducing visual bugs and a mobile game studio increasing weekly pull requests from approximately 50 to 1,900 using similar agent coordination.

Sources:The New Stack

Research

CodeRabbit Tests GPT-6.2 Astra: Captures Seven More Known Issues but Precision Drops to 12%

Quick take
2026-11-20 04:00 GMT+8

In code review, capturing more issues does not automatically translate to a better review experience. After testing GPT-6.2 Astra within its own pipeline, AI code review company CodeRabbit found that while the new model did identify a few additional known issues, the cost was a significant increase in comment volume and a sharp decline in precision—a combination precisely what review tools aim to avoid.

CodeRabbit compared Astra 6.2 against the previous generation Astra integration using a benchmark set of 83 known issues. Astra 6.2 caught 58 issues (69.9%), an increase of seven over the previous generation's 51 (61.5%). However, reported comments rose from 331 to 847, and precision fell from 21.74% to 12.22%—meaning only about one actionable comment per eight published corresponds to a known issue in the benchmark. The authors emphasize that these results do not establish a pure model improvement, as the baseline configuration utilized Sol 6.1 and other models, with each configuration run only once.

The problem with comment inflation is not merely numerical. Of the 589 standard actionable comments, only 72 matched known issues; the remaining 517 may contain genuine problems not annotated in the benchmark, but the authors explicitly state that understanding their value requires reading them line by line. CodeRabbit notes that every comment demands the author distinguish useful findings from noise. When a pull request is filled with low-value feedback, authors may eventually stop reading or addressing comments, causing useful findings to be ignored as well.

Sol 6.2 exhibited a similar but less pronounced pattern. In tests involving 81 known issues, Sol 6.2's high-reasoning configuration caught two more issues than 6.1 (74.1% versus 71.6%), but comment volume increased from 187 to 290, and precision dropped from 35.3% to 25.9%, with recorded costs rising by approximately 48%. Again, these runs cannot isolate the effect of reasoning levels, as different models were swapped into different parts of the pipeline.

Regarding code generation, the authors did not run matched benchmarks but shared practical coding experiences. Astra often produced a useful initial interface version from short prompts and, in long tasks, could return to the original goal after deviations. Sol 6.2 required more directional guidance; its initial UI designs showed no improvement over 6.1 and repeatedly provided incorrect configuration files, forcing the team to verify project startup capabilities one by one.

CodeRabbit's conclusion is to hold off on switching. Astra's additional catches came with a substantial increase in comments, and the integration still needs comparison against fixed routing configurations; Sol's marginal gains came with more comments and higher costs. The authors suggest reading incremental findings before switching and retesting the issues each configuration stopped catching. While OpenAI claims new models can complete demanding tasks with fewer output tokens, actual review costs depend on delegated calls and retry counts; lower token unit prices alone do not establish lower review costs.

Sources:CodeRabbit

Research

Google Announces 'Computer Use' Feature for Mac Gemini App, Enabling Direct Control of Desktop Software and Websites

Material
2026-11-20 10:20 GMT+8

On November 20, 2026, Google announced the addition of a new feature called "Computer Use" to its Mac version of the Gemini app. This functionality allows artificial intelligence to directly operate desktop applications and websites on the computer, performing actions such as clicking buttons, switching menus, entering text, and filling out forms on behalf of the user to complete complex tasks that span multiple software and web services.

A notable aspect of this update is that Gemini can perform these operations within an independent background window, displaying a real-time preview in the chat interface. This means users can continue using their Mac's main screen and mouse for other work while the AI handles tasks like data organization or form filling without interference.

Regarding security, Google allows users to decide which applications Gemini can operate and provides a one-click option to abort tasks. When the AI encounters sensitive procedures such as login or payment screens, the system automatically pauses execution and requires user confirmation before proceeding.

Google listed several application scenarios, including designers using AI to batch-adjust photo colors or set up 3D scenes, logistics personnel having AI extract information from local PDFs and automatically fill it into website quote forms, and creating presentations that compile data from multiple sources.

Currently, the feature is being rolled out to Google AI Plus, Pro, and Ultra subscribers, with initial support limited to US English. Users can enable it by adding @Computer use to their prompts.

Sources:INSIDE

Research

JEDEC Previews DDR6 Standard: Module Length Shrinks to 74mm, Bandwidth Increases Over 50%

Material
Verified 2026-11-20 10:40 GMT+8

The semiconductor standards body JEDEC released a preview of its next-generation DRAM standard, "DDR6," on November 18, 2026. The standard is currently under development, and the features shown reflect current plans that may change during the approval process.

In terms of performance, DDR6 RDIMMs will start at data rates of 12.8 GT/s, aiming for speeds exceeding 14.4 GT/s. The data bus width expands from 64 DQ in DDR5 to 96 DQ, a 50% increase. This architectural adjustment aims to deliver more than a 50% bandwidth increase at introduction through wider channels and higher frequencies.

For hardware designers, the most significant change lies in physical dimensions. The DDR6 module length is set at 74 mm, breaking the 133.35 mm (5.25-inch) standard that has persisted for about 30 years since the SDRAM era. This miniaturization not only improves memory density within systems but may also blur the lines between desktop and laptop modules, as 74 mm is only slightly larger than the current DDR5 SO-DIMM's 69.6 mm.

DDR6 adopts a pseudo-channel architecture to divide memory access and doubles the bank count to optimize refresh operations and data bus efficiency. Additionally, individual DRAM die densities will support up to 128 Gb, with further capacity gains achieved through optimized stacking within packages.

Regarding reliability and energy efficiency, DDR6 supports multiple system-level error-correcting code (ECC) implementations, allowing designers to balance cost, power, and reliability. The interface is optimized for high-capacity modules, explicitly targeting general-purpose computing as well as AI inference and agentic AI applications that demand greater capacity and performance close to the CPU.

Sources:jedec.org

Research

US House Panel Seeks Microsoft Briefing on Reported ByteDance Azure OpenAI Deal

Material
2026-11-20 09:58 GMT+8

The US House Select Committee on the Chinese Communist Party requested a briefing from Microsoft on November 18, following reports by technology media outlet The Information that the company is close to a $500 million deal allowing ByteDance to use OpenAI models through the Azure cloud platform.

Committee Chairman John Moolenaar, a Republican representative from Michigan, stated that Microsoft should not sell AI tools to ByteDance. He noted that US companies helping China catch up in the AI race is a serious concern, and that selling AI services to ByteDance is unacceptable given Microsoft's status as a Pentagon contractor.

According to reports, OpenAI itself restricts most Chinese enterprises from directly using its models, but Microsoft holds the rights to sell OpenAI technology and reportedly does not impose the same restrictions, giving ByteDance access to models it could not otherwise obtain via Azure. These details remain part of ongoing negotiations, and neither company has announced a completed transaction.

A Microsoft spokesperson responded that the company takes the misuse of AI services seriously and commits to enforcing terms of service and complying with applicable laws, but declined to comment on confidential customer negotiations. This dispute highlights a shift in Washington's regulatory focus from chip exports to cloud-based model services.

Sources:INSIDE

Research

European Commission Confirms Cyberattack Using AI Tool ARTEX; Some Internal Systems Still Down

Material
2026-11-20 09:53 GMT+8

The European Commission confirmed on Thursday (November 19) that it detected a cyberattack using AI tools on October 22. This marks the first time the EU's executive branch has publicly acknowledged being targeted by such automated offensive frameworks. As of now, some internal systems remain unavailable, and the extent of potential employee data leakage is still under investigation.

According to Politico and subsequent Commission statements, the attack affected several critical internal systems, including Sysper (the HR platform handling salaries, leave, and timesheets), MyPMO (medical expense reimbursement), and internal translation services. The attackers utilized ARTEX, an open-source AI penetration testing framework released by developer "Autumn-27" in July 2026. ARTEX uses large language models (LLMs) to autonomously execute vulnerability scanning, reconnaissance, exploitation, and privilege escalation.

The Federal Bureau of Investigation (FBI) issued a security alert on November 13 stating that ARTEX had been weaponized by threat actors since August, targeting sectors including finance, government, and technology. The FBI noted that some intrusions progressed from initial targeting to privileged access in under three hours. Although the original developer removed the GitHub repository on October 8, the code continues to proliferate. Investigations by Cloudflare and Spectrum Security indicate that ARTEX deployments and modified derivatives are increasing across the open web.

The EU cybersecurity agency CERT-EU had warned as early as November 5 that ARTEX might be used against EU institutions but did not name specific victims. The Commission's confirmation closes this gap. Spokesperson Thomas Regnier stated that while no evidence of mass data exfiltration has been found yet, investigators are verifying whether sensitive information such as names, salary grades, home addresses, and contact details were accessed. By November 19, Sysper had been restored, but the medical system JSIS remained offline.

Sources:iThome

Research

Bank of America Cuts MongoDB Price Target to $450, Citing Lower Atlas Growth Expectations

Material
2026-11-20 02:55 GMT+8

Bank of America has cut its price target for MongoDB Inc (NASDAQ: MDB) to $450 from $540 ahead of the company's third-quarter fiscal year 2027 earnings report, while maintaining a 'buy' rating. The bank noted that investor expectations for growth in MongoDB's Atlas cloud database business have declined, reflecting a broader drop in valuation multiples across the infrastructure software sector.

Analysts cited recent management changes as increasing uncertainty around sales execution and enterprise customer expansion, though they characterized these as execution risks rather than evidence of weakening competitiveness. A 29% year-over-year revenue growth rate for Atlas is now considered sufficient to meet investor expectations. Additionally, MongoDB's November 17 announcement of its acquisition of Firebolt aims to expand its capabilities into real-time online analytical processing, positioning it as a broader AI data platform competitor against Snowflake and Databricks.

Bank of America estimates the AI data platform market could reach $495 billion by 2030. The firm lowered its valuation multiple for MongoDB to 9.9 times estimated calendar 2027 revenue from 12 times. If fourth-quarter Atlas revenue guidance comes in at 26%, it would align with existing outlooks and support expected growth recovery; a weaker forecast could delay the investment case.

Sources:Proactiveinvestors UK

Research

Amundi Investment Institute: 2027 AI Investment Shifts from Rewarding Capacity to Demanding Returns, Holding Bottlenecks Rather Than Spending-Dependent Entities

Structural
2026-11-19 16:48 GMT+8

The Amundi Investment Institute, the research division of French asset manager Amundi S.A., published its 2027 AI investment outlook on November 19, 2026, authored by Federico Cesarini, Head of Developed Market Equity, FX, and Digital Assets, and Claire Huang, Senior Asia Macro Strategist. The article's core judgment is that AI has absorbed a disproportionate share of global capital, earnings growth, and investor attention, but the narrative is shifting from who can invest the most to who can capture the highest returns.

The financial basis for this shift is that free cash flow at U.S. hyperscale cloud service providers is expected to trough in 2027, with capital expenditure exceeding operating cash flow. The article argues that the market will care more about why free cash flow recovers than when it does, demanding early evidence that the monetization curve is bending upward. Any wavering in spending guidance or disappointing earnings could trigger a sharp re-rating of the entire AI infrastructure complex.

Building on this judgment, the article proposes a specific industry screening rule: limit exposure to industries requiring further commitments. Spending-dependent companies—particularly new cloud providers, data center developers, and colocation REITs—may face direct revenue impairment if spending plateaus. In contrast, true infrastructure participants (power equipment, electrical grids, networking equipment, chip fabrication) are more insulated, as supply is constrained, backlogs extend beyond 2027, and physical bottlenecks persist.

The article's most distinctive argument lies in its focus on the middle layer. The market typically frames AI as a two-horse race between enablers (building infrastructure) and adopters (using infrastructure), but this overlooks the facilitator layer—namely, IT services and software firms that translate AI capabilities into enterprise workflows. The article argues that the facilitator layer is structurally positioned to monetize earlier and faces a much lower bar of expectations. The transition to consumption-based pricing, the acceleration of agentic AI adoption, and the structural role of IT services as implementation partners all point to a monetization inflection point that is closer and less priced in than that of the infrastructure layer.

Regarding China, the article presents a risk perspective distinct from common frameworks: although Chinese AI companies carry small weight in global portfolios, their impact on the AI trade is far from marginal. Open-weight models compress frontier pricing power, parts of the technology stack are commoditizing, and efficiency gains achieved under compute scarcity question the capital expenditure intensity assumed by the U.S. build-out cycle. The article considers this the only risk among those discussed that challenges the returns thesis itself, rather than merely challenging its timing.

At the portfolio construction level, the article suggests placing greater emphasis on cash flow visibility, capital discipline, valuations, and the sustainability of semiconductor profits as AI investment shifts from capacity expansion to return generation. Dividend income and attractive valuations can provide complementary returns, particularly as market leadership broadens beyond the capital-intensive segments of the AI trade; this helps dampen volatility and provides a steadier return stream.

Sources:Amundi Research Center

Research

Omid Ashtari, Founder of UK AI Gaming Companion Kult: Routing Decisions by Confidence Level with Lightweight Models Answering First and Escalating When Uncertain

Material
2026-11-20

An AI gaming companion sitting next to a player must simultaneously satisfy three contradictory requirements: answers must be fast (games are real-time), accurate (each game has its own items, systems, and vocabulary), and low-cost (an always-on companion handles multiple turns per session, each of which must be cheap enough for a consumer product). Omid Ashtari, founder and CEO of London-based startup Kult, explained in a Google Cloud blog how they solved this trilemma using a model routing architecture.

The core idea is not to select a single model, but to let different tiers of models do their respective jobs. The lightest tier, Gemini Flash, handles high-frequency repetitive work—Kult initiates thousands of small calls for every game added. Gemini Flash runs every conversational turn: it determines what the player is asking and writes an answer based on Kult's self-built game knowledge base, fast enough to keep up with the pace of play. Meanwhile, the newest and heaviest Gemini models handle rare, deep reasoning tasks performed once per game. Ashtari noted that the latest Flash models have achieved a balance of price, speed, and quality that they could not find elsewhere, which is precisely why an always-on gaming companion can function as a viable consumer product.

However, tiering alone is insufficient; a mechanism is needed to decide when to upgrade. Kult uses a decision-specific model named Jev to handle such judgments: it does not write content, but returns a choice along with a calibrated confidence score. Cases with high confidence are resolved in a fraction of a second at minimal cost, while uncertain cases are escalated to larger Gemini models. Ashtari explained that much of the work in the system is inherently about decision-making rather than writing—it needs to understand which game the query concerns, whether the question requires a lookup, and whether the answer covers the key points. These decisions do not require large models, but they do need to be reliable.

For this routing to work, there is a prerequisite: answer accuracy does not come from the model's memory, but from Kult's own structured knowledge base. Games may change content with every patch update, and the training cutoff dates of general-purpose models mean they can be months out of date. Kult automatically builds structured knowledge bases for every covered game and keeps them updated; the model's job is to reason over this knowledge rather than rely on memory from training data. Ashtari's design separates factual sources from reasoning capabilities: the knowledge base provides facts, while the model handles reasoning and expression.

Kult currently covers more than 100 games, and adding new games is an automated process rather than a manual project. This allows their small team to target any game players pick up, including new releases—where most gameplay occurs in the first few weeks after launch. The company is currently in the alpha stage, gradually opening access to more players.

It is important to emphasize that this is Kult's self-reported architectural description. The Jev model is not public, so the quality of its confidence calibration cannot be independently verified; the product is in the alpha stage, and there is no public performance data proving that this routing strategy remains effective at scale. However, the design logic provided by Ashtari—using the smallest model capable of doing the job correctly, and using a decision model to determine when to escalate—offers reference value for any team building multi-turn conversational consumer AI products.

Sources:Google Cloud Blog

Research

Google Launches Tool Retrieval in Gemini API to Optimize Large-Scale Agent Performance

Material
2026-11-20 02:00 GMT+8

On November 19, 2026, Google announced the preview launch of "Tool Retrieval" within the Interactions API of the Gemini API. This update aims to address the context bloat problem faced by AI agents when connecting to dozens or even hundreds of tools. In traditional setups, developers must send the full parameter schema for every tool in each request, which inflates input token counts, adds latency, and lowers tool-selection accuracy as the candidate list grows.

By enabling Tool Retrieval, an agent can keep its full tool library on standby and load parameter schemas only when the current task requires them. Developers simply add a tool_search tool to their request and set defer_loading: true for the tools they wish to hold back. The model still sees the name and description of every deferred tool, so it knows what is available, but the detailed parameter definitions stay out of the initial context window. When the model decides to use a tool, the API loads its full schema within the same request cycle, allowing the model to emit the function call normally.

This mechanism requires no index building or hosting from the developer; the search runs server-side on Gemini's infrastructure. For teams maintaining their own tool registries or permission filters, Google also offers a client-side retrieval mode where the application executes the search logic, though this adds one round trip per search. Additionally, Tool Retrieval is compatible with remote MCP (Model Context Protocol) servers, allowing entire servers' toolsets to be deferred at once.

Regarding pricing, Tool Retrieval incurs no separate charge; input and output tokens are billed at standard rates. While enabling tool_search adds a small fixed block of input tokens, and loaded schemas continue to be billed for the rest of the conversation, the savings from never loading unused schemas typically outweigh these costs for large libraries. The feature currently supports Gemini 3.6 and later models, including all Gemini 4 series models.

Sources:Google

Research
Next →
Sign in / Sign up
ReadingResearchRadarInvestment framework

It must be emphasized that these are attributed arguments from the Amundi Investment Institute, not verified market outcomes. The article's judgments regarding the 2027 free cash flow trough and the IT sector accounting for 70% of net income growth in the MSCI ACWI are predictive in nature. The effectiveness of its framework depends on whether capital expenditure truly plateaus—if AI application acceleration drives a new spending cycle, the classification premise for spending-dependent companies will fail.