🎧 Listening note — language performance and understanding
In this Joy of Why conversation, Janna Levin asks Ellie Pavlick how language models represent language and what their human-like—and sometimes alien—behavior can tell us. Pavlick first separates the broad field of natural language processing from the particular neural-network approach of LLMs, then asks whether similar behavior implies similar mechanisms or meaning.
Pair it with lesson 1.1.1 as an open research discussion, not evidence that machine understanding has been settled.
Quanta Magazine
📘 1.1.1 — Behavior, understanding, and symbol grounding
Topic: Philosophy and History of AI
Three questions are easy to blur: Can a machine converse convincingly? Does its computation amount to understanding? How could its symbols connect to what they refer to? Keep them separate.
Turing’s move (1950) is methodological: replace “Can machines think?” with questions about performance in an imitation game. His original setup begins with a man, a woman and a separated interrogator; he then asks what happens when a machine takes one participant’s role. Written answers remove voice cues. The familiar textbook version asks whether text dialogue lets an interrogator distinguish a machine from a human. A behavioral criterion is not itself a construction recipe for intelligence.
Searle’s Chinese Room (1980) challenges a stronger inference. Imagine someone who knows no Chinese following English instructions to transform Chinese symbols by shape. Stipulate that the answers look fluent to outsiders. Searle argues that executing the program alone does not suffice for understanding or intentionality. This is a philosophical argument, not an experiment proving that current AI lacks minds.
The systems reply contests the level of attribution: perhaps the whole organized system understands even if its operator does not. Searle replies by imagining the operator memorizing the entire rulebook and database. The argument and reply remain disputed; neither should be taught as a settled empirical result.
Harnad’s symbol grounding problem (1990) asks how meanings could be intrinsic to a system rather than supplied by its interpreter. A dictionary in an entirely unknown language illustrates the difficulty: definitions point to more symbols. His candidate route starts with nonsymbolic sensory representations: iconic patterns support discrimination; categorical feature detectors support identifying kinds of things. Category names can then participate in symbolic descriptions and compositio… Read in full →
⚙️ llm-d 1.0 targets more reliable AI serving
IBM, Red Hat and Google-backed contributors have announced llm-d 1.0, an open-source, Kubernetes-native system for running AI inference across distributed infrastructure.
According to IBM, the release focuses on stable APIs, sharing resources among workloads with different priorities, and upgrades without interrupting service. It works with inference engines such as vLLM and SGLang rather than replacing the models themselves.
Why it matters: AI agents repeatedly revisit the same context. Routing requests toward servers that already cache it can avoid repeated computation. The underlying caching and distributed-serving techniques predate version 1.0; this release emphasizes their operational integration.
Production-readiness and cost benefits are the maintainers’ claims, not an independent guarantee for every deployment.
Published: 20 October 2026.
IBM Research
🔍 IBM aims to inspect AI models without disabling fast inference
IBM Research has introduced vLLM-Hook MIA, a plugin designed to capture and steer a language model’s internal states while preserving vLLM’s graph replay.
Many interpretability tools insert Python operations that interrupt this fast execution path. MIA instead prepares GPU operators before graph capture, then routes requests for internal states during replay.
The aim is to make runtime monitoring and interventions practical without retraining a model. These are IBM’s engineering claims; access to internal activations does not itself prove that a model is truthful or safe, and overhead depends on the model and intervention.
Published: 21 October 2026.
IBM Research
🔬 A microscope reveals the traffic inside living cells
Mitochondria move around slower chloroplasts in a turtle-vine leaf in a time-lapse video featured in Quanta’s selection from the 2026 Nikon Small World competitions.
The gallery explains confocal laser scanning microscopy: fluorescent labels mark structures, a laser illuminates a thin layer, and scanning builds an image one depth level at a time.
The images offer a view of cellular activity, rather than evidence that every proposed role of mitochondria has been established. Mycologist Patrick Hickey suggests that their movement may relate to signaling and interactions with other organelles.
Published: 21 October 2026.
Quanta Magazine
⚙️ AWS shows three routes to train Qwen3 on Trainium
A new AWS walkthrough uses torchtitan on a single trn2.48xlarge instance to train Qwen3 through three configurations: a Hugging Face Transformers v5 backend, torchtitan’s native implementation, and Neuron Kernel Interface kernels for mixture-of-experts layers.
The configurations share a launch pattern and configuration structure. The guide also demonstrates loading and saving Hugging Face checkpoints and registering a custom rotary-embedding kernel.
This is an implementation guide, not a new Qwen model release or a comparative performance benchmark. Its mixture-of-experts example reduces Qwen3-30B-A3B to four layers and 64 experts to fit the instance; it should not be mistaken for training the complete published architecture.
Published: 21 October 2026.
AWS Machine Learning Blog
📘 1.1.2 — Symbolic vs. connectionist AI
Topic: Foundations of artificial intelligence
Symbolic and connectionist AI differ in how they represent information and compute with it. A symbolic system can express knowledge as propositions and constraints. A connectionist system uses networks of numerical units with adjustable weights.
In a symbolic representation, the structure is explicit. Suppose a small scheduling problem states that two meetings cannot occupy the same room at the same time. Boolean variables can represent possible bookings; logical clauses rule out conflicting assignments. The textbook describes why this structure can make constraints concise, modular and easier to debug. Correct reasoning still depends on encoding the right constraints.
A neural network takes a different route. Stanford’s CS231n notes build a classifier from matrix multiplications separated by a nonlinear activation. Its weights are learned from data. The intermediate activations are numerical representations, rather than a supplied list of logical clauses. Without the nonlinearity, those stacked matrix multiplications collapse into a single linear transformation.
Learning those representations is the point of the classic 1986 back-propagation paper by Rumelhart, Hinton and Williams. Their procedure adjusts connection weights to reduce output error; hidden units come to represent features useful for the task. This describes a learning mechanism, not a guarantee that every hidden unit has a simple verbal meaning.
For comparison, imagine recognizing handwritten digits. A rule-based design could start with explicit tests on strokes; a neural design could learn intermediate features from labeled images. This is an illustrative contrast, not a benchmark result or a claim that either design always wins.
Keep the categories separate from claims about minds. Stanford explicitly warns that artificial neurons are very coarse models of biological neurons. A network’s numerical computatio… Read in full →
A reachable goal lies in a finite directed graph with unequal positive edge costs. Plain breadth-first search uses a FIFO frontier. What does its first returned solution guarantee?
Fewest edges, but not necessarily the lowest total cost
The lowest total cost, regardless of the number of edges
The smallest heuristic estimate among all goal paths
⚙️ AWS separates model and tool gateways in an agent-platform blueprint
AWS has published a reference architecture for a multi-account agent platform built around Amazon Bedrock AgentCore.
Its central design choice is to separate shared governance from team-owned execution. A platform team supplies approved templates and common controls, while individual workstream cells isolate runtime, memory and tool access. Separate inference and tool gateways govern model requests and tool calls.
The blueprint also includes evaluation gates, rollback and retirement checks. These are proposed engineering controls and reference implementation patterns—not independent evidence that an agent is reliable, secure or ready for a particular organization.
The useful question is not just whether an agent can call a tool, but who owns that tool, what the agent is allowed to access, and what evidence is required before a release reaches production.
Published: 28 October 2026.
AWS Machine Learning Blog
📘 1.1.3 — AI winters and milestones
Topic: Foundations of artificial intelligence
AI history is not a steady climb from primitive rules to inevitable intelligence. It is a sequence of useful ideas, impressive demonstrations and difficult tests of whether those demonstrations scale.
The 1956 Dartmouth workshop helped establish artificial intelligence as a research field. Early programs proved theorems and played games; these were genuine milestones. But success on a small, carefully structured task did not establish that the same method could handle an open-ended world.
One obstacle was combinatorial explosion. Searching through possible action sequences can work in a tiny puzzle and become impractical when the number of choices or the solution length grows. Faster hardware helps, but does not erase the underlying growth. The 1973 Lighthill report criticized this failure to scale and contributed to major reductions in British AI research support.
Another obstacle concerned representation. Minsky and Papert’s 1969 analysis exposed limits of simple perceptrons. A single linear threshold unit cannot express XOR: it cannot separate the two inputs that differ from the two that agree. That limitation does not prove that multilayer neural networks are incapable of learning; confusing a restricted model with an entire research direction is the wrong inference.
Expert systems offered another route: encode specialist knowledge rather than rely on general search alone. DENDRAL used chemistry knowledge to narrow candidate molecular structures. Commercial expert systems then expanded rapidly. Yet complex systems were expensive to build and maintain, and their handling of uncertainty and inability to learn from experience created serious weaknesses. After the late-1980s boom, many businesses failed to meet their promises—the period commonly called an AI winter.
The recovery involved several changes, not one magical invention. Neural-network research revived around backpropag… Read in full →
🆕 Gemini 4 Carbon — a restricted first rollout
Company-reported release facts and changes:
Google announced Gemini 4 Carbon on 28 October, building on Gemini 4 Argon. The company emphasizes longer engineering workflows, coding and cybersecurity defense. It reports improvements over Argon in secure-code generation and fewer unsupported claims in internal testing. These are company-reported findings, not guarantees for every deployment.
Access is initially limited: Google says it will begin with a subset of cyber defenders and expand to trusted partners through its Fairwind Program. Broader developer, business and consumer availability is planned rather than established by this announcement.
Artificial Analysis already lists Carbon (high) among its leading models. That does not establish a universal winner or an apples-to-apples improvement over Argon. The matched predecessor comparison below remains pending verification of exact scores and evaluation conditions.
Requested Artificial Analysis Intelligence Index matched Carbon–Argon comparison (v5.0; points); status as of 2026-10-29: independent results pending; no ranking.
Limitations: Pending means this exact matched predecessor comparison is not yet verified, NOT that independent evaluations do not exist. The retrieved page lists Carbon (high), but exact paired scores, model variants, test dates and full evaluation setup were not extracted. No invented scores, delta or universal ranking.
Google DeepMind — Gemini 4 Carbon
Artificial Analysis — model comparison
In a factory, 1% of parts are defective. A scanner flags 90% of defective parts and 10% of nondefective parts. A randomly selected part is flagged. What is the probability that it is defective?
A* has four frontier paths ending at A, B, C and D. Their (g, h) values are A=(2,8), B=(6,1), C=(4,4), D=(9,0), where g is cost so far and h estimates remaining cost. Which path does standard A* select next?
A regression model must predict the same real number c for three equally weighted training targets: 1, 2 and 9. Which c minimizes the mean squared error on these three examples?
🧠 Microsoft’s research roundup puts code understanding in focus
Microsoft Research’s 29 October roundup highlights a stubborn problem for coding agents: understanding unfamiliar code, not simply changing it.
The episode describes CABRA, a framework that generates coding challenges with controlled complexity to help isolate where agents struggle. Microsoft reports that tools helped agents overcome some programming obstacles, while understanding unfamiliar codebases remained difficult.
Another featured project, TeleTune, explores learning reusable computer-use skills from activity logs. Such logs record actions, but do not necessarily reveal the user’s goals or separate one task from another.
These are distinct research projects summarized by Microsoft—not a single new agent release or independent proof that the approaches will generalize to every real-world workflow. The useful distinction is between making an agent act and establishing that it understands the task it is acting on.
Microsoft Research
A CSP has D(X)={1,2,3,4}, D(Y)={2,3}, and the only constraint X<Y. After revising only the directed arc X→Y to make X arc-consistent with Y, what is D(X)?
Looking to get acquainted with neuroscience? Join a friendly space to learn the basics, share resources, and explore the field together.
👉 Join here: https://t.me/neuroprocessing_community
📩 Share this with anyone who might be interested.
🧬 AI searches human proteins for mRNA-delivery scaffolds
Researchers reporting in Nature Communications have developed DeepDelivery, an AI-guided platform for finding human proteins that could help carry mRNA into cells.
The system nominated 512 candidates from the human proteome. In experiments on 15 selected proteins, 12 supported particle formation and mRNA delivery within the researchers’ modular engineering framework. Several belonged to the TRIM family, better known for roles in antiviral responses.
The team also used model interpretation and perturbation analyses to guide changes to TRIM25, improving delivery in their experimental system.
The important boundary is cultured cells, not demonstrated treatment in people. These results identify promising scaffold components; they do not establish that a human-derived carrier is automatically safe, avoids immune reactions, or is ready for clinical use.
Nature Communications
Choose the action with maximum expected utility. A: utility 12 with probability 0.25, otherwise 0. B: utility 4 for certain. C: utility 10 with probability 0.5, otherwise −4. D: utility 30 with probability 0.1, otherwise −2. Which action is optimal?