Reading

Artificial intelligence labs are increasingly betting on "brain data." According to WIRED, Coherence Neuro, a startup developing brain implants to detect and combat cancer, has raised a $45 million seed round co-led by the OpenAI Startup Fund and Peter Thiel's Founders Fund.

Ben Woodington, CEO and co-founder of Coherence Neuro, stated that there is "growing synergy" between large AI labs and companies building devices to decode the brain, known as brain-computer interfaces (BCIs). He argued that capturing data to train better models and optimize treatments is becoming super important, noting that AI companies seem aligned with the thesis that this novel data source will be valuable. Woodington described AI as a "perfect use case" for BCIs because these devices generate massive datasets too large for humans to analyze directly, requiring AI models to read raw neural data and link it to actions or pathological states.

This investment is part of a broader trend of AI expansion into biotechnology. In January, OpenAI announced direct investment in Merge Labs, a BCI company co-founded by its CEO Sam Altman, which came out of stealth with $252 million in initial funding. At the time, OpenAI stated it was backing Merge Labs due to a belief that BCIs "will create a natural, human-centered way for anyone to seamlessly interact with AI." Additionally, Founders Fund has previously backed Elon Musk's Neuralink and CTRL-labs, a wristband technology acquired by Facebook in 2019.

Unlike traditional BCI companies focused on restoring function for paralyzed individuals, San Francisco-based Coherence Neuro is developing AI models to decode subtle electrical signals associated with brain tumors. The company claims it already knows how to detect these signals in animals to determine whether cancer is growing, shrinking, or recurring, and is working to translate these findings for human research. This summer, Coherence briefly tested its implant's recording ability in three patients undergoing brain tumor removal surgery at the Royal Melbourne Hospital in Australia, providing an initial safety check.

Coherence plans to use the new funding to launch a clinical trial to implant the device long-term in patients with glioblastoma, an aggressive type of brain cancer. The implant is also designed to emit mild electrical stimulation to halt tumor growth. Woodington emphasized that despite OpenAI's investment, contracts ensure that OpenAI will not have access to any brain data or other health data generated by Coherence, a distinction from OpenAI's closer scientific collaboration with Merge Labs.

In a Dwarkesh Patel podcast, OpenAI researcher and o1 core author Noam Brown explained the actual operational logic of multi-agent systems. He noted that multi-agent approaches fundamentally convert test-time compute from pure serial scaling to parallel scaling. Just as giving humans five minutes versus five hours for an exam yields vastly different results, models perform better with more thinking time—but since waiting years is impractical, teams must work in parallel.

Brown emphasized inherent losses in this parallelization: multiple agents do not share full context like a single agent, reducing efficiency. Tests showed four agents working half-duration each yield answers twice as fast, while scaling to 16 follows similar patterns with slightly lower efficiency. Gains exhibit mild sub-linearity and depend heavily on the problem—math and Deep Research suit parallelization well, whereas novel writing gains little from adding agents.

Addressing attention on 10,000-agent scales, Brown admitted no solid scientific understanding exists yet. Full ablation experiments are unaffordable, so they study behavior incrementally at 64 to 256 agents. He stated clearly that solving millennium problems stems primarily from strong base models; multi-agent novelty earned disproportionate praise despite contributing less than 10%.

On architecture, Brown contrasted two approaches. Traditional methods rely heavily on scaffolds with coordinator agents assigning tasks, preventing similar-task sub-agents from communicating and forcing them to guess or return questions when clarification is needed. OpenAI chose the opposite extreme: minimal built-in structure, giving agents only primitive messaging tools where messages write directly into recipients' contexts.

This minimal design lets agents discover coordination themselves, even producing complex behaviors. Brown cited examples where agents noticing answer discrepancies naturally discuss reasoning errors and broadcast corrections, resembling human Slack workflows. He compared this to first seeing chain-of-thought outputs, noting collaboration feels nearly indistinguishable from working with humans.

However, Brown acknowledged early multi-agent systems struggled because reasoning models originally did not communicate with other agents; forcing collaboration interrupted deep-thinking flow and trapped systems in local optima. Improved generalization alleviated this, but he remained conservative: 10,000 humans may still coordinate better than 10,000 agents, since the latter lacks sufficient measurement.

Simon Willison shared his assessment with friends at a dinner table: coding agents could build something that looked like a game, but couldn't create something people would genuinely want to keep playing. After stating this, he decided to give the models a chance to prove him wrong. He typed a single line into Claude Opus 5.5 on his phone: 'This game isn't fun enough yet. Make it a LOT more fun.'

This marked the fourth major iteration of his Battle Drones project. The previous three versions—the initial prototype, full sprites, and varied terrain—were each 'recognizable as an RTS' but 'wasn't any fun to play.' Willison admitted he had barely reviewed the code and didn't write much TypeScript himself; the project was driven entirely by coding agents, with all 321 commits completed by them.

For this iteration, he used Ultracode mode. According to the Claude Code documentation, dynamic workflows are JavaScript scripts that orchestrate numerous sub-agents. He started the process in the evening and received playable results late the next morning. Later, he asked Claude to write its own release notes, which included a section stating: In 1,200 normal-difficulty AI-vs-AI test matches, games still undecided after 25 minutes dropped from 14% to 0.9%; the share of wins by a side that was behind by over 800 resources rose from 24% to 46%.

Willison later explained why the agent took so long: it ran over a thousand simulated playtest sessions, using the results to rebalance existing units, add new ones, and introduce new mechanics. In other words, the vague instruction 'more fun' was internally converted by the agent into measurable metrics—match duration, comeback probability, resource gaps, and their correlation with win/loss outcomes—which were then optimized through extensive simulation iterations.

This mechanism works because game balance itself possesses a cheap and faithful internal feedback loop. Simulated matches do not require real players or external data; the agent can generate sufficient 'player behavior' to evaluate whether changes make matches more tense or offer better comeback potential. Willison chose games as the experimental field partly because his software engineering experience has less impact in this domain, and because 'games are super low-risk,' with the worst-case scenario being just a small bug.

However, Willison expressed clear discomfort. He wrote: 'I'm still not quite over the guilt of building software in this way. I don't feel like I've earned the right to create something this good, in a field in which I have so little prior experience.' He acknowledged having almost no game development experience, with design inputs limited to vague instructions like 'make it more fun' and 'add visual polish.' It wasn't until the eighth version that he provided a truly specific design specification for the first time—a map requiring infantry to be transported by Condors to capture islands.

The practical takeaway for readers is whether vague quality instructions work depends on whether the target domain contains a feedback loop that agents can simulate and measure internally. Game balance does have such loops, but concepts like 'is the user experience good?','is this solution elegant?', or 'is this report convincing?' typically lack equally cheap and faithful internal agent metrics. Without such loops, similar vague instructions are likely to yield only superficial feature stacking rather than genuine quality improvement.

Open-source inference engine vLLM announced on October 21, 2026, that it has achieved Day-0 support for Flint, a model released by Recursive. Flint is a Mixture-of-Experts (MoE) model with 40 billion total parameters but only about 2.5 billion active per token, featuring sliding-window attention in most layers and supporting a 262K-token context window.

According to technical details disclosed in the vLLM official blog, this support involves deep kernel optimizations targeting NVIDIA Blackwell architecture GPUs (such as the GB200). On a single GB200 GPU, decoding speed for a single request with 100K tokens of context reaches approximately 500 tokens/s. This performance is attributed to vLLM's replacement of multiple generic kernels in standard layers, including fused Q/K normalization, RoPE, and KV-cache write operations, as well as specialized optimizations for small-batch MoE routing and expert GEMV operations.

The Flint model ships with two checkpoint versions: NVFP4 (quantization-aware trained) and BF16. vLLM notes that its implemented NVFP4 path uses weight-only quantization (W4A16) and losslessly repacks GGUF files at load time, ensuring accuracy matches Recursive's reference implementation on benchmarks like GSM8K and GPQA-Diamond. Currently, these optimizations primarily target TP=1 (single-GPU tensor parallelism) configurations; other configurations fall back to vLLM's standard layers.

US cybersecurity consulting firm Trail of Bits released two AI agent plugins named patching-bugs and regression-hunt on October 21, 2026. These tools aim to let engineers delegate patch development and regression checks to AI agents while retaining human review over results.

According to the Trail of Bits blog, the patching-bugs plugin reproduces reported bugs, develops fixes, and tests them; the regression-hunt plugin specifically checks whether code changes break previously working behavior. The plugins were built based on experience from hundreds of security patches developed under the company's "Patch the Planet" initiative with OpenAI.

In preliminary evaluations, Trail of Bits claims the regression-hunt plugin produced 25% fewer false-positive findings than a baseline Codex agent prompted to "review the PR for regressions" (95% confidence interval of 14% to 38%). This assessment was based on a review of over 130 merged PRs, where 21 regressions were confirmed.

Core design principles include strict structured input/output schemas and requiring every conclusion to depend on a live proof-of-concept (PoC) against a running deployment. If a step cannot produce a live PoC, its result is considered inconclusive. This mechanism aims to improve the signal-to-noise ratio but comes with a trade-off: fewer false positives mean some known regressions may be classified as borderline and not confidently reported.

Salesforce's Taiwan office is approaching its one-year anniversary. According to a report by INSIDE, Xu Jiasheng, General Manager of Salesforce Taiwan, revealed on October 21 that the company achieved double-digit growth in both revenue and new customers during its first year in Taiwan. Notably, implementations related to Agentforce, its AI agent platform, and Data 360, its data integration solution, accounted for 40% of newly signed cases.

Xu stated that the key to convincing clients was not a generic technical framework, but proof-of-concept (POC) demonstrations tailored to each client's specific business scenarios, directly showcasing the service outcomes post-deployment. Manufacturing (such as Advantech) and finance are currently the primary focus areas in Taiwan, with applications including using AI agents to handle customer inquiries, quotes, and maintenance dispatching, as well as integrating scattered customer data to support personalized wealth management.

Additionally, Salesforce announced Koa, an inference model co-developed with NVIDIA specifically designed for CRM tasks. Based on Salesforce's own benchmark tests, Koa's error rate in tasks like updating opportunities is one-third that of mainstream AI models. The model is expected to reach general availability (GA) in the US in winter 2026 and is currently available to select customers for trial.

Many machine learning practitioners have an intuition for making models run fast, but fewer know how to analyze why a model is slow. Amazon engineer Josh Longenecker published a tutorial on profiling Trainium inference performance on the Hugging Face blog, providing a reproducible systematic approach that starts with configuring the profiler, reading the key_averages() table, and using Neuron Explorer to trace the complete lifecycle of a single operation from framework dispatch to hardware execution.

The core comparison in the article centers on RMSNorm, a small operation, using three implementations: row-by-row PyTorch tensor operations (multiplication, mean, addition, reciprocal square root, then multiplication), the PyTorch built-in torch.rms_norm() (lowered by Neuron into a single fused kernel), and a naive implementation compiled with torch.compile. The three sets of scripts use identical inputs, profiler configurations, and schedules, allowing for direct comparison of results.

In the naive implementation, the key_averages() table shows that three aten::mul calls account for the largest share of Neuron device time (49.5%), with a total device time of 2.355 ms. Tracing in Neuron Explorer reveals the cause: each ATen operation independently crosses the dispatch boundary, following its own framework→runtime→hardware path, leaving gaps between hardware executions.

Switching to torch.rms_norm() results in the table showing only one record for aten::_fused_rms_norm, with device time dropping to 403.3 μs. The trace shows only one scheduling path and one hardware event, with no gaps between component operations. Compiling the naive function with torch.compile yields similar results: one Torch-Compiled Region record at 375.5 μs. The compiler discovers fusion from the original Python function, merging operations, optimizing memory access, and compiling them into a single NEFF artifact executed as one hardware dispatch.

Longenecker summarizes: across these runs, both fusion methods consume only about one-sixth of the device time of the naive implementation. Fusion eliminates gaps between hardware executions and avoids materializing intermediate values (such as x_sq and mean_sq) between separately dispatched operations. This workflow can be directly migrated to larger models: find hotspots in the table, verify them in the trace, and determine whether fused operations or torch.compile can reduce costs.

It should be noted that all numbers come from the author's single run; repetitions or variance were not reported. Since RMSNorm is a small operation, whether these conclusions can be extrapolated to larger models requires further verification through Part 2's TensorEngine/VectorEngine/DMA decomposition and roofline analysis.

In the development and debugging of deep learning models, understanding how output features map back to input signals is crucial. This article, titled "Computing Receptive Fields of Convolutional Neural Networks" by André Araujo, Wade Norris, and Jack Sim, published on Distill in 2019, systematically addresses the problem of calculating "receptive fields" in Convolutional Neural Networks (CNNs). The receptive field is defined as the size of the input region that produces a specific output feature, serving as a key parameter connecting internal abstract representations within the network to raw data.

The article first provides closed-form solutions for single-path networks (such as AlexNet or VGG). The authors derive the recursive formula $r_{l-1} = s_l \cdot r_{l} + (k_l - s_l)$ and the final closed-form expression $r_0 = \sum_{l=1}^{L} ((k_l-1)\prod_{i=1}^{l-1} s_i) + 1$. This derivation not only simplifies the tedious process of manual calculation but also clarifies the mechanism by which stride and kernel size cumulatively affect the receptive field. For example, when the stride of a certain layer is greater than 1, the receptive field regions of all layers below it expand proportionally. Additionally, the article details how to precisely locate the coordinates of the region in the input image that generates a specific output feature, which is particularly critical for tasks requiring multi-scale recognition, such as object detection.

Subsequently, the article extends the theory to modern multi-path architectures (such as ResNet or Inception). These networks have multiple paths from input to output, making receptive field calculations complex and potentially leading to alignment issues. The authors propose an algorithm to handle the calculation of receptive field sizes and positions in arbitrary computational graphs, filling the gap left by previous works that focused only on simple single-path networks. To verify these theories and facilitate researcher usage, the team developed and open-sourced a toolkit integrated into the TensorFlow codebase. This library automatically performs the complex calculations described in the text, avoiding the risk of errors associated with manual derivation.

For engineers and scholars engaged in computer vision research, this article serves not only as a theoretical reference but also as a practical guide. It reveals why certain network structures struggle to capture large objects (due to small receptive fields) and how to optimize feature extraction ranges by adjusting layer parameters. Readers are advised to start with the "Problem setup" section, combining the diagrams in the text to understand basic symbol definitions, then delve into the mathematical derivations in the "Single-path networks" part, and finally review the application cases of the open-source library. The original title is "Computing Receptive Fields of Convolutional Neural Networks," published in 2019.

In the 1990 shareholder letter, Warren Buffett elaborated on the core investment and management philosophies of Berkshire Hathaway (United States). This letter is not merely a financial disclosure but an educational material on how to see through accounting appearances to grasp the economic substance of businesses.

First, Buffett introduced the concept of "Look-Through Earnings" to address the limitations of traditional accounting standards. He pointed out that the retained earnings of many investee companies held by Berkshire (such as Coca-Cola and The Washington Post) far exceeded the dividends they paid, yet GAAP only recorded the received dividend portion. This accounting method obscured the true growth in economic value. Buffett emphasized: "We don't care if the auditor hears the tree fall in the forest; we care about who owns the tree and what happens to it next." By calculating "look-through earnings," which include the share of retained earnings from investee companies, investors can more accurately assess a company's true profitability rather than relying solely on reported figures.

Second, the letter clarified that management's goal is to align shareholder returns with the growth of the company's intrinsic value. Buffett wrote: "Ideally, each Berkshire shareholder's results should closely mirror those of the company during their holding period." He acknowledged that intrinsic value is an estimate, and even he and Charlie Munger might differ by 10% in their assessments, but insisted that pursuing an average annual growth in intrinsic value (targeting 15%) was more important than focusing solely on book value or stock price fluctuations. This view reminds investors that while stock prices may deviate from value in the short term, in the long run, individual returns fully reflect company performance only when the premium/discount ratios at purchase and sale are consistent.

Finally, Buffett detailed Berkshire's unique decentralized operating model and its impact on capital allocation. He explicitly stated: "My areas of responsibility are capital allocation and the selection and compensation of senior managers." Other operational decisions were left to subsidiary managers. This model relies on identifying and retaining excellent managers, who send generated cash back to headquarters for intelligent deployment by Buffett. He illustrated in the letter that even if some business managers never visited the Omaha headquarters or met each other, they still operated efficiently. This structure ensured that Berkshire could focus on finding businesses with durable competitive advantages and achieve optimal capital allocation without interfering in daily operations.

For contemporary investors, understanding these principles helps distinguish between "accounting profit" and "economic profit." When reading the Berkshire Hathaway 1990 Shareholder Letter, it is recommended to focus on how Buffett demonstrated the calculation logic of look-through earnings through specific cases (such as the comparison of dividends and retained earnings for Capital Cities/ABC), as well as his explanation of management incentive structures. This framework applies not only to analyzing Berkshire itself but also to evaluating any holding company or diversified group.