标签：#tex

There's a school of voice-agent development that treats turn detection as something you buy, not something you build. Pick a streaming STT provider, let its end-of-turn logic decide when the user is done, and move on. For a lot of teams that's the right move — and if you're weighing the options, our breakdown of turn detection vs forced endpoints is the place to start. But some teams have already solved turn detection. They've tuned their own voice-activity detection over thousands of calls, they know their audio, and they trust their endpointing more than any default. For those teams, a streaming model's built-in turn logic isn't a feature — it's something to work around. What they want is narrower and faster: hand over a finished chunk of speech, get accurate text back, get out of the way. That's the case for bringing your own turn detection and pairing it with fast ASR over HTTP. Turn detection is an architectural decision, not a default Here's the framing that matters. In a streaming setup, the STT model is a participant in the conversation — it's watching the audio and deciding, continuously, whether the user has finished. That's genuinely useful when you want the provider to own that judgment. But it means the model is inserting its own decision between "user stopped talking" and "you get the transcript." If you already know the turn is over — because your VAD just fired — you don't want the model deliberating. You want it transcribing. Every millisecond the STT layer spends re-deciding a question you've already answered is latency you're adding for no benefit. So the decision isn't "which provider has the best turn detection." For these teams it's "who owns the turn boundary?" If the answer is you, then the ideal STT layer is one that does exactly one thing: turn a finished clip into accurate text, fast. Built-in vs. bring-your-own Built-in (streaming). The model reads tonality, pacing, and rhythm to detect end-of-turn — with Universal-3.5 Pro Realtime, aroun

You've got a recording and you want text back. For years that meant one thing at AssemblyAI: submit the file, wait for the job to finish, get a transcript. Async. It's reliable, it's cheap, and for a huge range of workloads it's exactly right. But "wait for the job to finish" is doing a lot of work in that sentence. If your file is two minutes long and your user is staring at a spinner, waiting is the whole problem. That's the gap the Sync API fills — and it's why "which transcription path" is no longer a two-way question. This post is about the two ways to transcribe a recording : async and sync. (If you're deciding between recorded and live audio in the first place — streaming versus the rest — start with our guide to real-time vs batch transcription , then come back here to choose between the two non-streaming paths.) The one-sentence difference Async transcription hands you a job: you submit audio, the work happens in the background, and you collect the result later by polling or via a webhook. Sync transcription hands you an answer: you POST a short clip and the transcript comes back in the same HTTP response — no job to track, no callback to wait for. Everything else follows from that. Async is built for throughput and depth on files of any length. Sync is built for speed on short files, when a person or an agent is waiting on the other end. How fast can each actually go? This is the question that usually settles it, so let's be concrete. Async processes the whole file and returns a single complete transcript, typically in seconds to a few minutes depending on file length and load. Crucially, it bills on audio duration ($0.21/hr on Universal-3.5 Pro), so a 30-minute file costs the same whether it comes back in 20 seconds or two minutes. You're optimizing for cost and completeness, not for the clock. Sync is built to return a transcript for a short clip almost immediately — roughly 134ms p50 — in one request/response, with no polling and no webhooks. It's price

"Context bombing" tricks hacking agents into shutting down before they can do harm.

DoorDash details the architecture behind Ask DoorDash, its AI-powered conversational shopping assistant, combining LLMs, specialized AI agents, MCP-based tooling, and an intelligence layer with persistent consumer memory and live backend data. Early results show up to 24% higher checkout conversion, 17% larger baskets, and improved intent accuracy using memory-backed sessions. By Leela Kumili

HyperTexting's new app aims to make the open web feel more like social media by turning websites, blogs, newsletters, and podcasts into a scrollable feed, while also making it easier to post to your own website.

GitHub has made the redesigned GitHub Copilot CLI terminal interface generally available. It adds a tabbed layout for sessions, gists, issues, and pull requests; an in-session, form-driven setup for MCP servers, skills, and plugins that avoids hand-editing config files; and a cleaner, theme-aware, more accessible UI with screen reader support. By Mark Silvester

Over the past week, two narratives have been colliding everywhere I look. On one side, there's panic. AI is expected to replace marketers, engineers, and entire categories of knowledge work almost overnight. On the other, there are quieter but far more consequential signals: enterprise teams discovering their AI infrastructure is burning through API budgets far faster than expected. This isn't because the underlying models are weak, but because the systems built around them are fundamentally inefficient by design. These aren't separate stories. They're the same failure showing up in different places. A conversation with another developer made that gap visible in real time. He argued that auditing a 150,000-line codebase requires feeding the entire repository into a model in one single, massive pass. It's still a common assumption in mainstream tech: that an LLM works like a giant biological brain that you must fully load with raw text before it can begin to think. But that assumption is already outdated. Modern AI systems don't scale through brute-force context. They scale through structure. And that shift changes everything. Key takeaways Bigger context windows did not solve AI. Treating a frontier model as a monolithic processor that re-reads an entire system on every query is wasteful, dilutes attention, and hides bugs under raw volume. ARC-AGI-3 makes the gap stark: frontier models scored under 1% on interactive reasoning tasks that untrained humans solve at nearly 100%. The gap is architecture, not memory. The teams pulling ahead treat the model as one narrow component inside a larger system: intelligent routing, task decomposition, retrieval, and only the minimum necessary context. The next advantage is not the biggest model or the longest prompt. It is the system designed around the model. Prompting was the first generation; systems architecture is the next. The Myth of the Infinite Context Window When context windows expanded into the hundreds of thousands o

LLM agents burn a ridiculous number of tokens on redundancy: opening the same files again and again, trying a patch, failing, then wandering back through the repo like they’ve never seen it before. A July 2026 paper, ContextSniper: AntTrail's Token-Efficient Code Memory for Repository-Level Program Repair , puts real numbers behind that waste. In repository-level repair, agents keep dragging in irrelevant code and logs. ContextSniper tackles that with a context layer built around tiered memory and an intention-aware context gate that filters low-value regions before they ever reach the model. That gate alone cut tokens by 51.5% on one host agent and 38.9% on Claude Code, while submitted-resolution rates stayed basically in the same neighborhood. The gate is the interesting part, because it is not tied to that paper’s exact system. It is a more general idea, and it is starting to show up across agent architectures. At heart, the gate is just a classifier. Given a request, it has to decide what kind of retrieval will answer the question cheapest: symbol lookup, semantic search, graph impact, mutation prep, or something else. That leads to the practical question the paper does not really answer: Do you need another LLM call just to decide what context to retrieve? We tested that directly. Five ways agents get code into context Before you can gate anything, you need a retrieval strategy. Most current systems fall into one of five rough families: Grounded read-only retrieval: parse the code and return exact symbol source by name. Byte-precise, no synthesis. Graph code intelligence: model calls, imports, entities, and dependencies as a graph, then traverse it. Embedding / RAG search: use vector similarity over chunks. Whole-repo packers: compress or dump the repo into the context window. Mutate / execute runtimes: retrieve context, then modify or run code. None of these is magic. Graphs are great for relationships, but they can drift away from source. RAG is useful, but f

Texas win at 5th Circuit left in place as attempts to overturn age law continue.

The Model Context Protocol team has promoted its Enterprise-Managed Authorisation extension to stable status, adding a centralised way for organisations to control access to MCP servers through their identity provider. The project states the aim is to replace per-server consent prompts with a zero-touch flow in which users sign in once and then access approved servers without further setup. By Matt Saunders

DeepSeek has previewed its V4 model family, led by a 1.6 trillion-parameter flagship, and made a one-million-token context window the default across all its services. The weights are downloadable and self-hostable, putting frontier-scale long context in reach of smaller labs and individuals without per-token payment to a closed provider. Key facts What: DeepSeek previewed two free-to-download V4 models that can read a million tokens at once, no longer as a premium add-on but as the standard setting. When: 2026-06-29 Primary source: read the source A large language model has no persistent memory. Each time it answers, it re-reads everything in front of it — your question, the conversation so far, any documents you pasted — and that pile of text is the context. The context window is the hard ceiling on how much it can hold at once. For years that ceiling was a few thousand words, then tens of thousands. Pushing it to a million has been possible but expensive, usually sold as a special, pricey tier. DeepSeek's move is to make a million the everyday default. The family comes in two sizes. V4-Pro is the big one — 1.6 trillion parameters in total, but only about 49 billion of them switch on for any given word. That design is called a mixture of experts : instead of running the entire brain for every token, the model routes each piece of text to a small relevant subset of specialists, so it stays affordable to run despite its enormous size. V4-Flash is the smaller, cheaper, faster sibling, meant for everyday chat and quick edits, and DeepSeek says it keeps up with Pro on simpler agent tasks. Making a million-token window affordable comes down to how the model handles its KV cache — the running set of notes it stores about every previous word, which grows steadily the longer the conversation gets. At a million tokens those notes become a mountain of memory, and the model normally has to consult every note for every new word it writes. DeepSeek's approach, which they call sp

If you run a RAG system in production, you eventually hit a dilemma that has nothing to do with your model and everything to do with your cache. Cache the answers to save tokens and latency, and one day a source document changes — but your cache keeps cheerfully serving the answer it built from the old document. Nobody gets an error. The number is just quietly wrong. Cache nothing , and every single call re-retrieves the same chunks, re-reads them, and re-pays the full context bill to rebuild an understanding you already built five minutes ago for a nearly identical question. Stale or expensive. Most teams pick "expensive" because at least it's correct, then bolt on a TTL and hope. This post is about why the TTL doesn't save you, and about two specific, mechanical fixes that let you cache RAG context and stay fresh. I maintain an open-source library called Coalent that implements both, so I'll use it for the runnable examples — but the two ideas are portable and worth stealing even if you never pip install anything. Failure mode 1: the stale RAG cache (and why a TTL won't save you) Here's the standard "answer cache" sitting in front of retrieval: answer = cache . get ( query ) if answer is None : chunks = retriever . retrieve ( query ) answer = llm . synthesize ( query , chunks ) cache . set ( query , answer , ttl = 3600 ) return answer This works until billing.md changes. The refund window goes from 30 days to 14. Your cache has an answer keyed on "what is our refund policy?" that says 30, and it will keep saying 30 for up to an hour — or forever, if the same question keeps refreshing a TTL that never expires under load. The reason this is hard is that the cache key (the query) has no relationship to the thing that changed (the source). You cached an answer; you threw away the fact that this particular answer was derived from billing.md . So when billing.md changes, you have no way to find the answers that depended on it. The TTL is a confession that you can't answ

X has launched a hosted MCP server, making it easier for developers to connect AI applications with the company’s API.

Elastic open-sourced Atlas, a system built on Elasticsearch that maintains three categories of memory for agents. Atlas integrates with agents via MCP and maintains per-user isolation of memories. When evaluated on question-answering capability, it scored 0.89 Recall@10. By Anthony Alford

The Knowledge Atom: Writing for Machines That Read The Hoarder's Reflex Everyone is learning to feed the machine. Bigger context files. Paste the whole document. "Give the AI all the context it needs." The entire industry has converged on a single instinct: when in doubt, add more. It's the wrong instinct. A context window is not a hard drive. It's a desk. And a desk piled with every document you own is not a well-informed desk — it's an unusable one. The model doesn't read better because you gave it more. It reads worse, because the one line that mattered is now buried under a thousand that didn't. Knowledge an AI can't find is knowledge it doesn't have. Knowledge it always carries is weight it always pays. The Two Failures There are only two ways to get this wrong, and almost everyone commits one of them. The first is the dump . You take everything you know and pour it inline — into the system prompt, the master config, the one document to rule them all. It feels thorough. It is the opposite. Every token you add dilutes every token already there. Signal drowns in completeness. The model now has all the knowledge and none of the focus. The second is the orphan . You did the disciplined thing. You wrote a clean, perfect note, in its own file, out of the way. And then nothing pointed to it. No index, no trigger, no path back. The note is immaculate and invisible — which is worse than never writing it, because you believe the knowledge is in the system when in fact it is dead. Both failures share one root: confusing having knowledge with retrieving it. Same Pattern, New Sauce Watch the field long enough and you'll see the same thing return, repainted each time. The "Ralph Wiggum" loop becomes "the agentic loop." Agent teams that talk to each other become a single orchestrator, and then an agent that makes other agents talk to each other. Every cycle sells itself as the breakthrough. Every cycle is a re-skin of the last. Underneath the churn, only one thing actually ch

Vercel has released Eve, an open-source framework for building, deploying, and operating AI agents in production. The framework uses a filesystem-based project structure to organize agent instructions, tools, skills, subagents, communication channels, and scheduled tasks, enabling developers to define agent behavior while reducing the amount of supporting infrastructure they need to implement. By Daniel Dominguez

Overleaf is an open-source, collaborative LaTeX editor that bundles MongoDB, Redis, and the ShareLaTeX application into a single self-hosted stack. This guide deploys Overleaf Community Edition using the official Toolkit plus a Traefik override that adds automatic HTTPS via Let's Encrypt. By the end, you'll have Overleaf serving collaborative LaTeX editing securely at your domain. Set Up the Project Directory 1. Clone the Overleaf Toolkit: $ git clone https://github.com/overleaf/toolkit.git ~/overleaf-toolkit $ cd ~/overleaf-toolkit 2. Initialize the configuration: $ bin/init This creates config/overleaf.rc , config/variables.env , and config/version . 3. Create a directory for Traefik file-provider routes: $ mkdir traefik-routes Configure Overleaf for a Reverse Proxy 1. Edit config/variables.env : $ nano config/variables.env OVERLEAF_APP_NAME = "My Overleaf Instance" OVERLEAF_SITE_URL = https://overleaf.example.com OVERLEAF_NAV_TITLE = "Overleaf CE" OVERLEAF_BEHIND_PROXY = true OVERLEAF_SECURE_COOKIE = true The last two flags are required when Overleaf is fronted by an HTTPS reverse proxy. 2. Edit config/overleaf.rc : $ nano config/overleaf.rc SIBLING_CONTAINERS_ENABLED = false 3. Create the project-level .env file used by the Traefik Compose file: $ nano .env DOMAIN = overleaf.example.com LETSENCRYPT_EMAIL = admin@example.com SERVER_IP = 192.0.2.1 SERVER_IP should be the server's public IP (Traefik binds 80/443 to it). Add the Traefik Route 1. Create the dynamic route: $ nano traefik-routes/overleaf.yml http : routers : overleaf : rule : " Host(`overleaf.example.com`)" service : overleaf entryPoints : - websecure tls : certResolver : le services : overleaf : loadBalancer : servers : - url : " http://sharelatex:80" 2. Create the Traefik Compose file: $ nano docker-compose.traefik.yml services : traefik : image : traefik:v3.6 container_name : traefik restart : unless-stopped command : - " --providers.docker=true" - " --providers.docker.exposedbydefault=false" - " --

I've been working as Software Developer mainly focussed on Java and builts many application using Eclipse RCP framework or VS Code Application. Almost all the time I had to deal with multiple large files (either read/generate/validate) them which seemed very difficult and some of them almost impossible as most of them would be dependant on each other and would be referencing each other (just like how java files work together). Now assume client1 requires the same content in multiple Json files and client2 needs it in xml files. We couldn't go on writing a different application or go on adding if conditions and blah blah blah !!!! Wouldn't it be easier if as soon as I execute the application it generates the content in whatever format I choose and also taking care of dependencies/ references (like adding import statements). Additionally integrate with features of IDE and provide proposals, perform validations on the fly. Rela World Examples : Try googling Arxml once (Trust me I've dealing with these files for almost 7 years and it's always a nightmare to debug these) Solution: Xtext framework In this tutorial, I will show you how to use Eclipse Xtext and Xtend to build a simple, readable DSL that automatically generates Java boilerplate for you. Fair Warning: There will be no running executions screenshots or anything. You are gonna have to run it yourself and check the results and of course questions are always welcome in the comments section. But if for some reason you are unable to replicate this then let me know I'll try to explain further. I believe the best way to learn is by doing it yourself. The Goal: What are we building? Instead of writing 100 lines of Java with private fields, getters, and setters, we want our developers to write 5 lines of code in our own custom language (basically you can create your own programming language with your own custom syntax), like this: entity User { var name : String var age : Integer } When this file (assume file extension

This is an adapted English version of an article I first wrote in Japanese. I work with AI to shape and review my drafts, but the argument and the field observations are my own. The numbers are cited from public surveys (linked at the end). I built an aggressive prompt-injection block to stop my AI agent from repeating the same mistakes. It worked, so I kept adding rules. By the time I noticed, the file had ballooned to 56,000 characters — and the agent had quietly stopped functioning. Too much context, attention spread too thin to act on any of it. I gutted it back to under 1,200 characters, and here's the part that still stings: it behaved better with fewer rules. That was the day I learned my whole mental model was backwards. This isn't a post about making your AI more accurate. It's about designing so that accuracy stops being the thing you depend on. The mistake I made for months My agent kept skipping the same step in a workflow. So I did what every engineer does on instinct: I added a rule. "Don't skip this step." Then it did something else dumb, so I added another rule. Then another. I was treating the rules file like a conversation with a colleague — as if the agent would remember yesterday's correction and carry it forward. It doesn't. Every run starts cold. "Be careful next time" assumes a next time that shares state with this time. For a stateless model, there is no continuity to appeal to. You are talking to a counterparty with no memory of the conversation you think you're having. So the rules pile up, because each correction feels like progress. And for a while the numbers even improve. But adding rules has a ceiling, and I blew straight through it: at 56,000 characters the agent wasn't reasoning over my guardrails anymore — it was drowning in them. Knowing a rule and stopping at it are different things Here's the distinction that took me far too long to see. Putting a rule in the context window means the model knows the rule. It does not mean the mod

Most agents drag their entire past into every turn. A better default: keep a thin index of what was said hot, and fetch only the few turns you actually need — intact, on demand. Code: github.com/NirajPandey05/jit_context There is a quiet assumption baked into how most agents handle memory: that more context is safer than less. If the model might need something, put it in the window. The conversation grows, every prior turn rides along on every new request, and we trust the model to find the part that matters. That assumption breaks twice. It breaks on cost , because an agent loop re-sends its whole window on every step — a hundred stale turns aren't paid for once, they're paid for on turn 101, 102, and every step after. And it breaks on quality , because models don't read a long window evenly. Relevant facts buried in the middle get underweighted; irrelevant bulk competes for attention with the thing that actually answers the question. Past a point, a bigger context produces a worse answer, not just a costlier one. So the interesting question isn't "how do we fit more in?" It's "how do we keep the window small and dense without losing the one old turn that matters?" This post is the design we built around that question — for the specific case of long conversation history — plus the benchmark we used to keep ourselves honest. 01 · The mechanism: a hot index over a cold store The design borrows directly from how computers have always managed memory that doesn't fit: a small fast tier that's always present, a large slow tier that holds the bulk, and a rule for moving things between them. Virtual memory pages between RAM and disk. We page between the context window and an external store — for attention instead of address space. Concretely, there are two tiers. The cold store holds every turn at full fidelity, keyed by id — nothing is thrown away. The hot index holds one compact entry per turn: a short summary, a little metadata (entities, whether the turn recorded a dec