标签：#AI

Have you ever been coding and then gotten into that flow state? You know where hours pass by , and it feels to you it's only ben a few minutes? Me too. One thing that really helps me get into that state is music. So I create my own music Lo-Fi server called compile and chill. As a part of this project, I created three radio stations. Each station can generate a 16:9 scene with Amazon Bedrock , compose an instrumental loop with ElevenLabs, and turn an illustration into a six-second video through OpenRouter. Generated files live in private Amazon S3 storage and return to the browser through the Nuxt server. I also added a Stream Deck API interface! This tutorial shows how to build this radio station from start to finish. The complete source code is available in the Compile & Chill repository . Watch the full video on YouTube . Prerequisites You need the following tools for the complete build: Node.js 22.19 or newer. The locked Nuxt 4.5.2 release requires Node 22.19+, 24.11+, or 26+. npm 10 or newer. An AWS account and a configured AWS Command Line Interface (AWS CLI) profile. The AWS Serverless Application Model (AWS SAM) CLI for the private storage stack. Access to Stability AI Stable Image Ultra through Amazon Bedrock in us-west-2 . An ElevenLabs API key for music generation. An OpenRouter API key for animated scenes. The provider credentials are optional. Without them, the UI, bundled scene, station switching, player, and Focus Block timer still work. The identity running the app needs bedrock:InvokeModel plus bucket-scoped permissions for s3:GetObject , s3:PutObject , s3:DeleteObject , s3:DeleteObjectVersion , and s3:ListBucketVersions . Use a role or profile scoped to the station bucket rather than an administrator identity. For this project I included infrastructure as code with SAM to help setup the AWS parts. It's also included in the repo. Steps 1. Run the station without credentials Pull down the repo and get started! git clone https://github.com/ErikCH/comp

The $200 million extension comes just months after the physical AI startup reached a $2 billion valuation.

When someone asks whether to use Azure OpenAI Service or the direct OpenAI API, the starting point is this: the models running on both platforms are identical. GPT-4o, GPT-5, and the o-series models you deploy on Azure have the same weights, the same capabilities, and the same output quality as the ones you call from platform.openai.com, and what changes between the two platforms is the infrastructure where they run, the authentication mechanism, and the compliance guarantees the provider can offer on those requests. What changed in 2026 Azure AI Foundry was renamed Microsoft Foundry on January 1, 2026, and Azure OpenAI Service now lives inside that unified platform alongside the model catalog, development tooling, and agents. References to Microsoft Foundry in new documentation point to what used to be Azure AI Foundry. In July 2026, the GPT-5.6 family arrived with Sol, Terra, and Luna available on Azure the same day as on the direct OpenAI API. Historically Azure lagged four to eight weeks behind new model releases because Microsoft validates them within their compliance frameworks before making them available, and while that gap still exists for some specific features and APIs, for the main models in the GPT-5 family availability is converging. Where data is processed When you call GPT-4o from the OpenAI API, the request goes to OpenAI's own infrastructure, which is centralized and gives you no control over which region processes your data. For most use cases that doesn't matter, but for organizations with data residency requirements, regulatory compliance needs, or industries like healthcare, banking, or government, that detail can determine whether the service is usable at all. Azure OpenAI runs the same models within the boundary of your Azure tenant, so the data you send in prompts doesn't leave to OpenAI's infrastructure but processes in the Azure regions you choose. That's what makes it possible to meet HIPAA, SOC 2, EU data residency, and other certificati

The False Dichotomy The agent ecosystem is split into two camps: Camp Loops (Boris Cherny, OpenAI Agents SDK, LangGraph): > "Agents are loops. Plan → act → observe → repeat. The loop is the atomic unit." Camp Graphs (Steve Yegge, Gas Town, LangGraph DAGs, CrewAI): > "Agents are graphs. Nodes are agents/tools. Edges are handoffs. The graph is the architecture." Both are right. Both are incomplete. What Loops Get Right Loops capture temporal behavior — the iterative, self-correcting nature of agent work: - Replanning on failure (AdaPlanner, ReAct) - Budget enforcement (token caps, step limits, cost ceilings) - Verification gates (process reward models, extraction floors) - Learning loops (feedback → lessons → advisory → suppress) A loop is a control structure. It says: keep going until condition X. What Graphs Get Right Graphs capture structural composition — how capabilities connect: - Handoffs (peer-to-peer control transfer) - Parallel execution (swarms, polecats, fan-out/fan-in) - Supervision trees (Erlang/OTP-style restart strategies) - Provenance (who called whom, with what context) A graph is a dependency structure. It says: A feeds B, B feeds C, C can restart A. The Missing Layer: A Compiler Between Repos and Runtime Here's what neither camp addresses: Where do the nodes come from? Today: - You find a repo on GitHub - You hope it implements what it claims - You wire it into your graph/loop - You pray it works There's no verification layer. No SBOM. No attestation. No provenance. HURCULES: The Compiler Between Repos and Runtime HURCULES sits between the repository and the agent runtime: GitHub Repository → HURCULES → Verified Capability Package → Agent Runtime (Loop or Graph) It doesn't care if your runtime is a loop or a graph. It produces verified capabilities that work in either. What HURCULES Compiles | Input | Output | |-------------------------------|---------------------------------------------------| | Raw repo (any language) | Deterministic map (file tr

Foundation models are expensive. A trading agent that calls GPT-4 for every price prediction burns budget fast. Lightweight forecasters are cheap to run but expensive to train, especially when you only have a handful of examples. MetaCaster introduces a meta-harness architecture where agents don't forecast directly. Instead, they train specialized lightweight models on-demand from few-shot examples and textual context. This is not another AutoML wrapper. The meta-agent orchestrates data generation, architecture selection, and training loops to produce task-specific forecasters in minutes. The result is a deployable model that runs inference without touching the foundation layer again. The Economic Gap Time-series forecasting in production faces a resource trap: Foundation models (TimeGPT, Chronos) deliver strong zero-shot performance but cost $0.002 to $0.02 per prediction at scale. Lightweight forecasters (PatchTST, DLinear, FEDformer) run for pennies but need thousands of training samples and hours of GPU time. Few-shot scenarios (new trading pairs, emerging markets, privacy-sensitive health data) don't have enough history to train from scratch. MetaCaster targets the intersection: resource-constrained environments where you need specialized models but can't afford foundation API calls or long training cycles. Meta-Harness Architecture The system has three layers: 1. Meta-Agent Orchestrator The top-level agent receives a few-shot time series (as few as 5-10 examples) and optional textual context (domain descriptions, seasonality hints). It decides: Which lightweight forecaster architecture to instantiate (PatchTST, DLinear, Autoformer, etc.) What synthetic data generation strategy to apply How to configure the training harness (learning rate, epochs, augmentation) The meta-agent uses a learned policy, not heuristics. It's pre-trained on a meta-dataset of diverse forecasting tasks so it generalizes to new domains. 2. Data Generation Agents These agents expand the f

📝 Originally published (in Japanese) at forge.workstyle.tech . When building voice-based AI interactions in the browser (avatars, voice bots, streaming AI), you’ll inevitably hit pitfalls stemming from audio physics and browser implementation quirks. This article compiles 16 traps I encountered during product development , organized in a symptom → cause → solution lookup format . No need to read from top to bottom—jump straight to the symptom you’re facing. Echo and Self-Response Issues 1. Avatar Responds to Its Own Voice (Despite echoCancellation: true ) Symptom : TTS audio is picked up by the mic, and STT recognizes it as user speech, creating a self-response loop. Cause : AEC (Acoustic Echo Cancellation) requires a reference signal (the "sound to cancel"). Only the browser's official playback paths ( <audio> / WebRTC receiver tracks) serve as references. Custom playback via Web Audio API does not reliably function as a reference . Solution : Return TTS audio from the server as a WebRTC remote track and play it via an <audio> element. This eliminates echoes without text-matching workarounds (tested: 99 seconds of continuous speech with speakers on, zero false user turn detections). 2. Echoes Are Gone, but Speaking Simultaneously with the Avatar Distorts My Voice and Causes Misrecognition Symptom : Only during dual speech, proper nouns get mangled (e.g., "社員数" → "シャインズ"), especially at word beginnings. Cause : Fundamental AEC trade-off. To cancel echoes, AEC suppresses/distorts near-end (user) audio during dual speech. Solution : Mitigate in three layers: ① Increase mic Opus bitrate and enable FEC (see Pitfall 13) ② Provide vocabulary hints to STT (see separate article: use "recent avatar speech" as initial_prompt , not a dictionary) ③ Instruct LLM: "Input is STT transcription with potential errors. Interpret unnatural words as phonetically similar terms and add confirmation prompts." 3. Can’t Suppress Audio from Other Apps (Music, Videos) Symptom : Audio/lyrics fr

Before Malone left, OpenAI had already reshuffled its infrastructure org, shifting his reporting line away from President Greg Brockman and putting Vice President Sachin Katti in charge of the group.

GitHub Copilot premium requests are the metered unit that determined how much advanced Copilot usage your plan covered, and if you are searching for how they work in mid-2026, you need two answers, not one. First, the mechanics: a premium request is consumed each time you use an advanced Copilot feature, scaled by a per-model multiplier, against a fixed monthly allowance that came with your plan. Second, the news: as of June 1, 2026, GitHub moved Copilot from request-based billing to usage-based billing , and premium requests are now officially labeled "legacy" throughout GitHub's own documentation. Their replacement is GitHub AI Credits, metered at one cent per credit. Both systems matter today. Annual Copilot Pro and Pro+ subscribers who stayed on their existing plans are still billed in premium requests, and every question about the new credits model (allowances, overages, admin controls) is easier to answer if you understand the system it replaced. Here is the complete picture, with the numbers. What is a premium request? GitHub's definition is simple: a request is any interaction where you ask Copilot to do something, whether that is generating code, answering a question, or reviewing a pull request. Routine interactions, like inline code completions, are unlimited on every paid plan and never touch the meter. Premium requests are the interactions that use more advanced processing, and they draw down a monthly allowance: Copilot Chat : one premium request per user prompt, multiplied by the model's rate (ask, edit, agent, and plan modes all count). Copilot code review : each review consumed one request originally; since June 1, 2026 it carries a 13x multiplier , so a single review deducts 13 premium requests. Copilot coding agent and CLI : one premium request per prompt or session, times the model's rate. Only your prompts count; the autonomous tool calls Copilot makes along the way do not. Spark : a fixed rate of four premium requests per prompt. The critical n

Following the discussion on named AI personas and trust — here's the engineering side: how do you keep AI-status disclosure genuinely persistent throughout a conversation without making the interface feel robotic or constantly interrupting the experience a named persona is meant to create? The Naive Approaches Both Fail Option A: One disclaimer, message one, never again. Trivially easy to implement, but gets forgotten within a few exchanges — exactly the failure mode worth avoiding for personas carrying real emotional weight. Option B: Repeat "I am an AI" every single message. Technically persistent, but breaks the actual UX a named persona is trying to create, and users will tune it out as noise within a few messages anyway — repetition without variation loses its signal value fast. Neither is a good engineering solution. The better pattern is contextual, adaptive disclosure. Pattern: Risk-Weighted Disclosure Frequency python class DisclosureManager: def init (self, base_interval=8, high_risk_interval=3): self.base_interval = base_interval self.high_risk_interval = high_risk_interval self.messages_since_disclosure = 0 def should_inject_disclosure(self, message_risk_level: str) -> bool: interval = ( self.high_risk_interval if message_risk_level == "high" else self.base_interval ) self.messages_since_disclosure += 1 if self.messages_since_disclosure >= interval: self.messages_since_disclosure = 0 return True return False message_risk_level comes from the same classification pass used for scope/escalation detection covered in earlier persona-guardrail architecture — emotionally sensitive or high-stakes exchanges trigger disclosure more frequently than routine ones. Pattern: Disclosure Woven Into Persona Voice, Not Bolted On Rather than an interrupting system message, integrate the reminder into the persona's actual response style: python def inject_natural_disclosure(response_text, persona_config): disclosure_phrases = persona_config.disclosure_variants # e.g. for "Ок

I want to describe a completely ordinary hour from my week, because I suspect it's your week too. ...

These are the lessons we learned evaluating LLMs for real-world secret scanning. The post How to evaluate LLMs before production appeared first on The GitHub Blog .

I want to be upfront about something. I didn't figure this out proactively. I figured it out after my second burnout in three years — sitting in a period of forced recovery, unable to look at a code editor without feeling a specific kind of dread that I couldn't logic my way out of. I'd done everything the burnout recovery advice said to do. Took time off. Set better boundaries at the new job. Worked on the psychological stuff. All of it helped. None of it explained why recovery felt so much harder and slower than it should. Then I got bloodwork done. And the picture became considerably less mysterious. The Diagnostic Output bash $ bloodwork --full-micronutrient-panel --date=recovery-period [CRITICAL] vitamin-d: 18 ng/mL target: 40-60 ng/mL status: severely deficient duration: estimated 2+ years note: dopamine synthesis impaired at this level [CRITICAL] rbc-magnesium: low note: serum looked normal — wrong metric duration: unknown — never previously tested correctly note: HPA axis running unregulated [HIGH] omega3-index: 3.1% target: 8%+ status: neuroinflammation elevated note: western diet + zero supplementation [HIGH] hs-crp: 2.9 mg/L target: <1.0 mg/L status: significant systemic inflammation note: never measured, thoroughly normalized [WARNING] ferritin: low-normal note: passing standard panel, causing fatigue bugs-found: 5 bugs-known: 0 recovery-speed: severely impaired by all of the above Two burnouts. Same underlying biology. Neither time did anyone suggest checking any of these markers. What the Numbers Actually Meant Vitamin D at 18 ng/mL: Vitamin D is a direct input to dopamine synthesis. The enzyme that produces dopamine requires it. I had been trying to rebuild motivation and find meaning in work — the core challenge of burnout recovery — while running a dopamine system without adequate substrate. javascript // what I was trying to do dopamine.rebuild() // what the system had to work with vitaminD: 18 // severely deficient tyrosineHydroxylase.efficiency:

Google has expanded Gemini in Chrome with a desktop workflow that lets users send a selected part of a web page directly to Gemini. Called Select from screen , the feature is designed for moments when a full page is not the relevant context: a user can draw a box around particular text, an image, or a mixed section of page content and ask Gemini to analyze or act on it in Chrome's side panel. The change makes Gemini more closely embedded in everyday browser work. Rather than manually describing what is on a page or switching between tools, users can identify the exact on-screen material they want Gemini to consider. For teams that regularly research products, review creative assets, compare information, or work from web-based documents, that can make AI assistance more immediate. Its usefulness will still depend on whether Gemini in Chrome is enabled for the user and, for managed environments, how administrators configure access. How Select from screen works Google's official instructions for sharing specific parts of a screen with Gemini in Chrome describe a straightforward process. Users open the Gemini side panel in Chrome, choose Select from screen , then draw around the area they want to share. The chosen content is sent to Gemini as the basis for the next interaction. The important distinction is that the feature is not limited to a single content type. Google says the selected region can contain text and/or images . That gives users a more precise way to supply context from a web page without treating the entire page as the prompt. Workflow element General Gemini interaction in Chrome Select from screen Context provided User supplies a request in the Chrome side panel User selects a defined region of a web page for Gemini Content types Depends on the interaction and context available Selected text, images, or a region containing both Selection method No region-selection step Draw a bounding box around the relevant content Why the workflow matters The value is

Variant Multiplier already let an editor swap one section of a winning ad and keep the rest. The next request from a real production job — replacing product SL-603 with SL-808, a different hearing-aid SKU, across an entire finished ad — was a different shape of problem. It's not "change one section," it's "change every mention of the product, everywhere it appears, while keeping literally everything else the same." Two direct quotes from the editor drove the whole five-PR arc: the transcript editing was too rigid for word-by-word changes, and separately, "the music, voice, etc. should retain the same, we should keep the quality the same, and not make it do a lot of changes." If a re-render can degrade something the editor explicitly asked to keep untouched, the render path is wrong for the job — no matter how good the model is. The cheap fix first: let editors actually edit PR #67 shipped before any product-swap work started, because it was the cheap, high-value half of the same feedback: "I am just able to select word by word here but I am not really able to change the whole sentence a lot easier," and separately, "I'm able to double click on these words and then just type it in." Both were UI gaps in the transcript editor, not pipeline gaps — selecting by sentence or scene instead of only by word, and retyping a line verbatim instead of only substituting individual words. Shipping this first, standalone, meant the harder product-swap work that followed didn't also have to carry an unrelated UX fix in its diff. A product catalog the tool never had PR #69, stacked directly on top of the transcript work, is pure groundwork with no user-visible feature of its own: a product catalog, because Variant Multiplier had no concept of "a product" at all before this. The editor's own framing made the requirement explicit: "have a product selection right here, for Pro Bluetooth, for [the other SKU], and maybe other tons of products" going forward. The catalog data itself is mai

Every AI video pipeline eventually has to answer an unglamorous question: what did we actually pay for that clip? On the main video-generation service, the answer for months had been "a hardcoded constant." That's fine until the vendor changes its own pricing, or a code path pays for the same synthesis twice, or a voice engine mints a clone, bills for it, and never sends it downstream. Over a ten-PR run I audited and rebuilt the voice and lip-sync pipeline from the billing layer up, then used the vendor's own SKU tiers to cut cost 7x without touching output quality. A cost model built from hardcoded constants isn't a cost model. It's a guess that happens to compile. Billing what the vendor actually charges PR #224 was workstream one of three from a sibling-tool audit: port the cost-accounting fixes that Presenter Generation and Variant Multiplier had already found, verifying each one against this repo's own code rather than assuming the same defect existed in the same place. Anthropic returns exact token counts on every response. Nothing in the pipeline read them — every charge was a hardcoded per-call constant, so the ledger and the vendor invoice diverged the moment usage drifted from whatever number had been typed in at launch. The same PR closed a second gap: two editor-facing routes could spend money — kicking off a generation, retrying a step — outside any run . A run is the unit everything else (budgets, audit trail, the cost ledger) is keyed to. A spend with no run attached is a spend the ledger can't even see, which is worse than a wrong number. Paying twice for a take the model returns unchanged PR #225 found the sibling bug's twin: some vendor calls return the exact same asset on a retry — no new synthesis happened — and the pipeline billed a second time anyway because "call succeeded" and "call did new work" were treated as the same fact. The fix is the boring, correct kind: hash the output, and only charge when the hash changes from the take you already

On the AI video ad platform I work on, every scene goes through the same painful loop: write a prompt, send it to an AI video model provider, wait two minutes, open the result, squint at the frame, and decide what went wrong. Camera too wide. Product missing from the hero shot. Color palette drifted warm when the brand brief says cool neutrals. Avatar looks like a different person than scene three. That loop was manual, slow, and expensive. Each regeneration burns GPU credits. Operators were becoming prompt engineers by accident — and still missing subtle failures until stitch time, when fixing scene four means re-rendering everything downstream. The insight behind vision-in-the-loop prompt authoring is simple: the model that wrote the prompt can also look at its own output and rewrite the prompt with surgical fixes. Not a full replan — a per-scene correction grounded in the actual generated frame, not the operator's memory of what they hoped would appear. The manual loop we were trying to kill Before this work shipped, the swipe iteration flow looked like this: Plan — Claude generates a scene-by-scene script with visual prompts Generate — each scene renders independently through an AI video model provider Review — operator opens the portal, compares frames to the reference ad Rewrite — operator edits prompts in a text field, often guessing at what the model misread Regenerate — repeat until acceptable or budget exhausted Steps three and four are where throughput dies. An experienced operator can spot "product not visible" in three seconds, but translating that into prompt language — "medium close-up, product centered in lower third, shallow depth of field" — takes another minute per scene. Multiply by twelve scenes and three swipe iterations, and a single ad creative consumes an hour of human attention that should be spent on brand strategy, not frame inspection. The generated frame is ground truth. The original prompt is a hypothesis. Vision-in-the-loop closes the

Meanwhile, paid accounts can configure prompts to trigger when things happen in Gmail, Slack or GitHub.

Record-breaking robot races are less substantial than household chore challenges.

CD Projekt Red is remastering The Witcher 3: Wild Hunt, the hit RPG that first launched in 2015, and the updated game will be available starting September 29th. The Witcher 3: Wild Hunt - Remastered will bring improvements to visuals and combat, a "reimagined" skill tree, and more, and it also includes the first two […]

The company's new fundraising total now stands at $232 million.