AI Trends Weekly · Airing

AI Trends Weekly 003 · OpenAI’s mathematics: progress and resistance

Mon, 12 Oct 2026 16:00:00 GMT

OpenAI released 700-plus AI mathematics manuscripts. Explore the research, corrections and objections, and how people and AI might collaborate.

2026-10-04 – 2026-10-10 · Airing

AI released 700-plus manuscripts. Why are mathematicians divided?
On October 6, OpenAI released mathematical research generated by an internal model. Some saw an acceleration of discovery; others called for a boycott. What lies behind these different responses?
Why can an agent forget a request from two hours ago?
Two Pi engineering articles examine limited working context and the cost of repeatedly processing conversation history.
Can an agent resume after crashing halfway through a task?
Pi Durable persists work for long-running agents. The difficult case arises when an external action has happened but its result has not yet been recorded.
Dream-RSI: using past experiments to plan the next attempt
The researchers turn code-optimization history into a replay environment, then use AI to revise the search policy. What does this “dreaming” save?
PANDA: how should a group of agents form a useful team?
This paper studies collaboration infrastructure: finding suitable partners, organizing communication around a task and continuing when a participant fails.
REA
Connect a coding agent to application and binary inspection through MCP, tracing how a feature works.
Designeer
A browsable directory of design tools, components, learning resources and design engineers.
GitAgent
Keep an agent’s identity, rules, memory and skills in a Git repository, with versioned changes.
Monid
A shared entry point for agents to discover and call tools through skills, MCP or a CLI.
OpenBot
Give desktop agents separate instructions and workspaces, with task handoff and shared files.
Rill
Bring browsing and Claude Code or Codex tasks into one window, connecting pages with project work.
Mintlify
Help people and agents read, search and use product documentation alongside API workflows.
Lapis One
A dedicated Debian computer for agents, with remote access and hardware connections to other devices.
vPhone Workstation
Manage virtual iPhones in a native macOS window, with a device library, settings and lifecycle controls.
T3 Code
An open-source graphical workspace for coding-agent threads, diffs and Git collaboration.
Turn recurring mistakes into failing checks
The /correct proposal turns recurring coding mistakes into lint rules, type checks or tests. Even when an agent forgets a verbal correction, the environment can report the same concrete failure. This preserves feedback where the failure condition can be stated clearly.
Names change; project constraints still need clarity
This reading discusses CLAUDE.md, AGENTS.md and customization of the environment around an agent. These files serve as project handoffs: how to build, how to validate and which constraints matter. Useful instructions make work startable; filenames and loading behavior remain tool-specific.
EmbeddingGemma 2: one retrieval space across modalities
Google’s model card describes text, image, video and audio embeddings in a shared 768-dimensional space, with selectively loaded encoders. Try a personal-library task, such as finding an image or recording from text, then evaluate retrieval and device overhead.
The Jev application list keeps growing
The saved inspection note records awesome-jev growing from 583 to 591 entries, including bisect, moderation and security work. Examine where each decision layer connects and what executes afterward. The count is a snapshot, not a maturity rating.
DEX #366: what computer does a personal agent need?
DEX groups personal agents, cloud desktops and AI computers. Compare where a task lives: a portable device, an always-on host or a temporary cloud desktop. Start with persistence, data access and recovery needs before choosing hardware or services.
Across 16 products, memory and initiative recur
Founder Park’s roundup spans personal agents, specialist assistants and memory tools. Use it as a candidate list and try one adjacent to a current task. Useful initiative depends on what a tool remembers, when it interrupts and whether you can inspect and correct it.
Thin Is In: where does control live in a thinner interface?
Thin Is In connects persistent agents with lighter interfaces. Work that once required opening an app and navigating several controls may begin with a short request. That creates a design question: where do progress, source material and opportunities to take over remain visible?
AIGC Weekly #191: a partial reading
The cached portion covers personal agents, AI decompilation and cross-device work. The stream reports roughly 21% coverage, so this note retains the excerpt’s threads without presenting a verdict on the entire newsletter. The full issue belongs in its subscription channel.
Does more AI usage mean better value?
Kazike compares two A16Z reports on usage frequency, payment and concentration of spending. The useful shift is from model capability to how people use AI. Users, paying users and the general population are different denominators. Spending indicates demand; completed tasks, time saved and continued use help explain the value delivered.
The gap between fingers: seeing Zeno’s paradox
The original post shows a child a finger gap repeatedly halving. Explanation need not begin with complicated notation. Our geometric-series sketch below makes one point: infinitely many remaining steps can have a finite sum; infinitely many steps alone do not imply infinite time.
Hands-On Large Language Models: learn through diagrams
Jay Alammar and Maarten Grootendorst’s Hands-On Large Language Models connects model concepts to applications through illustrations and code. It provides useful background for this issue: tokens, vector retrieval and generation conditioned on earlier input. Understanding the mechanism makes it easier to locate what a product claim actually changes.

AI Trends Weekly 002 · Opus 5.5: making films with code

Sun, 04 Oct 2026 16:00:00 GMT

From films built with code to leaner agents: how do faster, cheaper tools become workflows we can control?

2026-09-27 – 2026-10-03 · Airing

Opus 5.5: making films with code
Opus writes animation, a browser renders frames, and editing tools assemble the film. The interesting change is how precisely a film can be revised.
Stamp-book motion: finish one small gesture
Page turn, stamp, next page. Study the rhythm and material transitions as UI or brand-motion inspiration; the example does not need a grand AI thesis.
Art history in 15 seconds: expression and compression
The creator’s Fable 5.5 clip offers visual transitions and stylistic continuity to study. Condensing history is a creative interpretation; a clip alone cannot establish historical accuracy, production time or reproducibility.
Faster technology still needs creative judgment
Katzenberg brings the discussion back to imagination, emotion and compensation for creators. Alongside the cover story, ask whether easier production also improves storytelling, choices and collaboration terms.
Less context, with evidence intact
Pi separates the archive from the working context; SoL-Pi reduces repetition. A smaller bill can also come with fewer solved tasks.
Pi Agent 1.0: two paths for agent work
Pi 1.0 refines the minimal terminal agent; the experimental Pi Durable explores long-running work. The release brings both choices into focus.
Claude Mods: tools that fit your workflow
From context displays to timely reminders, the terminal agent is becoming a customizable work environment.
From writing code to operating systems
The Tencent Cloud SRE case connects infrastructure, observations and knowledge in a graph, and stages capability from read-only diagnosis to approved recommendations and execution. The lesson is gradual delegation: demonstrate explanations before permitting operational changes.
Define the work before assembling an agent team
The Opus team example is a starting point. Decide role boundaries, handoff artifacts and acceptance checks first; independently verifiable work keeps collaboration from merely adding conversation.
When does Jev routing actually save money?
Five Jev saves explore division of labor: do classification, scoring and routing always require full generative reasoning?
90.5% at about a fifth of the cost: an eval case
A 44-ticket support case shows why a higher score needs a named dataset—and an account of what the agent changed.
Bring in an advisor at decision points
Use the main model for execution and an advisor before a plan, after repeated failures or before completion. This is a community workflow idea; confirm current configuration and model availability rather than using two models for every step.
ElevenLabs
A hosted platform for narration, dubbing and conversational voice.
VoiceStudio
A community-recommended local tool for cloning, synthesis and dubbing.
drawgent
Connect coding agents to a shared Excalidraw canvas.
Chit
Turn local Claude Code history into project-based receipts.
Hemory
Convert everyday speech into searchable memory for agents.
NOAN
A shared source of business facts and context for agents.
Rene
A personal assistant spanning mobile and shared workflows.
Invent a Dataset
Generate task-specific training data from an intent.
Maestro
Describe mobile and web UI tests with readable flows.
e2e · TesterArmy
Combine deterministic locators and AI steps in one test.
Sonnet 5.5: efficiency for everyday work
The official launch emphasizes speed and efficiency. Revisit model choice for bounded tasks such as fixes and documents; similar benchmark scores do not mean interchangeable performance on every workload.
DevDay and Jev: decisions as a platform capability
The two saves summarize DevDay and debate the Decisions API’s impact on Jev. They point toward decisions becoming a platform capability; “killing” a competitor is the author’s interpretation. Compare availability, output guarantees, latency, cost and switching overhead.
DEX #365: what remains scarce when motion is easy?
This email issue connects ubiquitous motion design with personal agents: lower production barriers do not automatically produce good work, and assistants still need resources and good interactions. It adds an editorial counterpoint to the cover. The saved email has no public original URL.
AIGC #190: an ecosystem snapshot
The accessible saved portion discusses Opus, Muse and creative examples. This issue indexes only that portion and does not treat the uncaptured paid remainder as read. Alongside DEX, it contrasts new releases with questions about experience and quality.
What does a personal assistant actually save?
“Nobody wants an AI assistant” is a provocative opinion, not demand data. The useful question is whether users want less mental effort without surrendering decisions. Start with specific repeated tasks and separate recommendations, approvals and execution.
After creation, think about discovery
This Xiaohongshu marketing save sits at the edge of the AI theme: production also needs discovery and conversion. It remains a short reading note. Timing, account strategy and traffic outcomes are the author’s experience, not guaranteed platform rules.
Behind the “AI ban” headline: terminology
The executive order concerns using SI instead of AI in applicable executive-branch communications; existing historical documents need not all be revised. The order does not itself ban the technology. Separate terminology, legal definitions and substantive measures before assessing impact.

AI Trends Weekly 001 · Jev: how should an agent make decisions?

Sun, 27 Sep 2026 16:00:00 GMT

Inside Jev’s decisions, the next steps in AI creation, Claude’s biology research, and products for everyday life.

2026-09-20 – 2026-09-26 · Airing

No long answer. How does Jev make a decision?
Four distinctions matter: a predefined answer space, parallel decisions, probability calibration, and business rules that remain the program’s responsibility.
Pi × Jev：Putting execution authority back into code
Routing, tool gates and answer verification each address a small question. The model supplies signals; the harness decides the consequences.
Jev and FLock: align the measurement conditions first
A community test of a thousand financial complaints raises useful questions. Model versions, the distribution of long inputs, network time and cost definitions determine where its conclusions can apply.
Jev’s limitations: turning them into workflow design
Parallelism pays off only when the task structure suits it. Subtitle selection, date extraction and Chinese-language evaluation reveal the respective boundaries of semantic judgment, deterministic computation and quality assurance.
Four kinds of Jev applications: decisions in real workflows
Browser, document, advertising and sales demos show different integration patterns. Public code reveals more than demo speed about exactly which step Jev performs.
Tencent’s debugging agent: recommendations need evidence
Moving from hours to minutes compresses the work of assembling evidence across systems. Read the 88% action-agreement rate alongside the sample’s purpose and its confidence rules.
A one-person Grok business: count review time and failures
$18K is a revenue target for five clients; the article does not show payment records. Its routing and acceptance design is useful, but operating it requires complete cost accounting and a way to handle exceptions.
Animate the character. Keep the ability to edit it.
What makes Rege’s example worth saving is the .coa project supplied with the demo. Inside it, expressions, hair movement and timing become separate places to make changes.
A pixel animation in five minutes: what comes after the first result?
Rikuo’s short clip is a distinctive style reference. Turning it into your own production capability requires visual specifications, editable deliverables and revision tests together.
463 video examples: turning them into reusable workflows
The reusable assets are character constraints, storyboard structure and failure records. The post, the repository on the save date and the current repository have different counts. Versions and counting units need to be explicit.
Fish Audio：What can you actually control in a voice?
Fish Audio’s emotion tags let scripts specify how a line is read. A complete voice workflow still needs reference voices, pronunciation of names, transitions between passages and accounting for repeated generations.
ScreenKite：How conversational editing meets the timeline
Recording, transcription, editing and B-roll become useful together when every instruction maps to inspectable clips and time ranges. Model costs, asset uploads and editable exports still require separate scrutiny.
Claude found ART. What is actually new?
AGI Hunt’s report offers a lead more interesting than a model expressing excitement: how much biological understanding advances when a known enzyme, neighboring genes and repeated RNA elements are connected?
After 21 hours: how to review Claude’s research
AI Era’s 950 agents makes a striking headline. Understanding the actual progress requires separating session counts, candidate counts, report counts and experimental results.
The value of frontline QA: knowing where things break
A secondhand account of an internal post cannot establish a company’s staffing decisions, but it raises a concrete question: as automation gets cheaper, who defines the real risks worth testing?
Four small products, four everyday obstacles
Children’s sense of remaining time, adults’ distraction, outdoor exploration and restarting habits all concern time and action, but each needs its own interaction mechanism.