We cooked using Opus 5.5.. This footer was designed & developed in literally one shot.

Meta open sourced the hardware for its Muse AI agent on October 2nd. By October 4th someone had forked it to talk to a completely different, fully open model instead. Muse Gadgets ships open firmware and a Linux SDK so hobbyists can build their own buttons, screens, and sensors on Raspberry Pi or ESP32 boards, Meta even gave away 5,000 free "Muse Home Link" units to seed it. What stayed closed is the brain. The device talks to Meta's cloud, Meta's models, Meta's account system, full stop. Within two days, a fork called Hermes Gadget stripped out the Muse branding, kept the ESP32 firmware and the round AMOLED display wiring, and pointed the whole stack at Nous Research's open Hermes Agent instead, self hosted, no cloud account, no vendor telemetry. "Hold a button. Ask Hermes. Hear the answer." Apache 2.0, running on a gateway you control. Teknium reposted it with three words: we got ESP32 at home. This is the actual shape of open source hardware in 2026. Companies open the parts that cost them nothing, the firmware, the enclosure, the wiring diagram, because that layer was never the product. The model is the product, and that's exactly the layer that stays locked. Developers noticed instantly, which is why the fork took 48 hours instead of months. I'd watch what gets forked fastest, not what gets announced loudest. That's the real signal of where the actual value sits, and right now it's sitting with whoever controls the model, not whoever ships the enclosure. x.lingyaoai.com/Teknium/status/2106632…

"How did they even know?" That's the question PewDiePie asked after OpenAI banned his account for the second time while he was building his own model. It's also the scariest sentence in AI right now, and nobody involved can actually answer it. The model is Ajax, a fine-tuned Qwen 3.5 9B model built to run locally inside his self-hosted Odysseus workspace, handling web search, browsing, and email without sending anything to a frontier API. One of the two bans was explicitly for distillation, training on another model's outputs. OpenAI's enforcement caught it. It never explained how, and PewDiePie says it never will. This isn't an isolated incident. OpenAI spent July disrupting a coordinated extraction campaign tied to Moonshot AI associates, over 4,000 fraudulent accounts running adversarial distillation against its models before the plug got pulled, so the detection and enforcement machinery is actively running at scale against funded operators, not just flagging stray home users. And Nadella has been publicly calling this out as what he calls a reverse information paradox: frontier labs ban distillation from their outputs while training on public data and on everyone's usage patterns, correction, and ratings. His framing is blunt, you pay for the model twice, once with the subscription and once with the data exhaust that trains the next version. Nobody is arguing OpenAI doesn't have the right to protect its weights. The actual problem is that the rule is enforced by a black box against a YouTuber with 109 million subscribers and, separately, against a rival lab's 4,000-account extraction operation, with the same opacity both times. A policy that can't tell you which line you crossed isn't a policy, it's a trapdoor. Sam Altman reacted to the video. He didn't answer the question either. x.lingyaoai.com/interesting_aIl/status…

A sandbox founder just declared that Browserbase is dead, pointing to session pricing at a fraction of the cost. Whether that specific number survives contact with reality or not, the direction is real. Browser infrastructure for agents turned into its own crowded category almost overnight, Browserbase raised at a nine figure valuation, Cloudflare shipped Browser Rendering on its own edge network, and a wave of cheaper VM and container based alternatives showed up arguing the whole category is overpriced for what it does. What gets lost in the pricing fight is that renting a headless Chrome instance was never the hard part of web automation. The hard part is what happens after the page loads. Sites change their markup constantly, and most scraping setups either break on every redesign or burn an LLM call re-reading the full page on every single run, which is slow and adds up fast at any real volume. I built DeepScrape around a different split. An LLM looks at a page once, works out the extraction logic, and generates CSS selectors that get reused deterministically after that, so a redesign triggers a quick re-heal instead of a full re-think every time. It runs on Playwright, has a /map endpoint for discovering URLs across a site before you touch them, supports autonomous agent navigation for multi-step flows, and exposes the whole thing as an MCP server so Claude or Cursor can call it directly as a tool instead of you gluing together a scraper on the side. Cheap browser compute solves infrastructure cost. It doesn't solve the part where your extraction breaks every time a site ships a frontend update. github.com/stretchcloud/deep… x.lingyaoai.com/AniC_dev/status/210660…

MCP gateways are quietly turning into the most important piece of enterprise agent infrastructure nobody outside platform teams has noticed yet, and Uber's internal setup is a good look at why. The pattern is a control plane paired with a proxy gateway: a registry that knows about every internal service, and a gateway that translates MCP calls into whatever the service actually speaks, gRPC, Thrift, plain HTTP. New services get auto-discovered from existing API definitions instead of someone hand-writing a wrapper for each one. Every tool is disabled by default until a team explicitly turns it on, and every call gets checked against who's allowed to use it and scrubbed for anything that shouldn't leave the service boundary. That's a very different shape from how most companies are still doing this, which is a developer standing up a one-off MCP server for whatever internal API they're touching that week. Block, Cloudflare, and Microsoft have all published versions of the same idea over the past year: don't expose services to agents one at a time, put a gateway in front of all of them and let the gateway handle auth, logging, and rate limits once. The tradeoff nobody wants to say out loud is that the gateway becomes the thing you have to get right. A bug in one hand-built MCP server breaks one integration. A bug in the gateway, or a permission that defaults the wrong way, touches every agent talking to every service behind it at once. Getting an agent to call an internal API was the easy part this year. Deciding, at scale, which agents get to call which APIs under what identity is turning out to be the actual work. x.lingyaoai.com/santtiagom_/status/210…

Europe finally has a frontier-scale open model worth talking about, and it came out of Heidelberg, not Paris or London. Aleph Alpha's Kolibri uses a mixture-of-experts setup: 384 experts total, only 6 active per token, 78.1 billion parameters on paper but just 3.46 billion doing the work on any given pass. That's the same trick DeepSeek and Mistral have leaned on to get frontier-ish performance without frontier-ish inference bills, applied by a company whose entire pitch has been sovereignty and government contracts rather than chatbot market share. The wider pattern here is more interesting than any one model. Mistral raised at a multi-billion euro valuation on the same sovereignty argument. The EU is funding its own compute clusters through the AI Gigafactories initiative. Germany's SAP and SAP-adjacent vendors keep signing deals that specify data residency before they specify benchmark scores. None of this is about beating GPT or Gemini on a leaderboard. It's about having a model a ministry of defense or a regulated bank can run without a US company's terms of service in the loop. What's easy to miss is that this only works commercially if the model is actually competitive, not just available. A sovereign model nobody wants to use because it's two generations behind is a procurement checkbox, not a product. Kolibri's MoE efficiency matters because it gives Aleph Alpha a shot at being both compliant and good enough, which is the combination that actually wins contracts instead of just satisfying auditors. So instead of one race to the frontier, there's a compliance lane running next to it, and Europe is finally building the cars for it instead of just writing the regulations. x.lingyaoai.com/Aleph__Alpha/status/21…

Every database GUI built before 2024 has the same problem now. It was designed for a human typing SQL by hand, and increasingly the thing querying through it is an agent that was never supposed to see a password prompt in the first place. Gridex is the latest answer: one native app for macOS, Windows and Linux that connects to Postgres, MySQL, SQLite, Redis, MongoDB, SQL Server and ClickHouse, built by a solo developer out of Vietnam. The Community edition is free with no account wall. The part worth paying attention to is the Pro tier's MCP server: thirteen tools split across three permission levels, SQL sanitization before anything executes, row count estimates before a query runs, and a full audit log of every call an agent makes against your data. It's not alone in trying to solve this. Google open sourced its own MCP Toolbox for Databases this year, a connector layer for the same problem at the infrastructure level rather than the GUI level. Beekeeper Studio and DBeaver are retrofitting AI panels onto interfaces that predate the agent era by half a decade. TablePlus still doesn't have a public MCP story. The gap between an AI assistant bolted onto an existing tool and a product built assuming an agent is a first class user from day one is where Gridex and Google's toolbox both sit, and it's a different design decision than most of the incumbents made. What actually matters here isn't the query editor or the ER diagrams, every competitor already has those. It's the permission model. The moment you give an agent a live database connection, your security boundary stops being the login screen and starts being whatever the MCP server decides to allow, tier by tier, call by call. Most teams haven't thought about that yet. The tools shipping now are the ones that have. x.lingyaoai.com/tom_doerr/status/21063…

A no-code web scraper Chrome extension just went up for sale for five thousand dollars. Seven months, $6,467 in total revenue, 817 active installs, a 4.2 rating, a featured badge on the Chrome Web Store, and the person who built it is done with it. That's the small end of a market where the large end just raised $75 million. Firecrawl closed a Series B led by Smash Ventures with YC, Altos Ventures, Nexus Venture Partners, Freestyle and Offline Ventures all in, and launched Alexandria the same week: a cloud dataset bundling tens of millions of scientific paper abstracts and matching code documentation so agents don't have to scrape it themselves. Same underlying problem at two completely different price points: getting clean web data in front of a model without burning a fortune doing it. Here's the part that doesn't show up in either story. Most scraping tools, no-code or funded, still make the LLM re-read the page on every single pass to figure out what to extract. That's the expensive part, not the request itself. Selector generation should happen once, get cached, and only call the model again when the page actually changes shape. That's the whole bet behind DeepScrape. Playwright drives the browser, an LLM generates the CSS selectors the first time it sees a page, and after that extraction runs on the cached selectors with no model call at all, self-healing only when the DOM shifts. It ships an MCP server so Claude and Cursor can call the scraper directly as a tool, plus a site-to-MCP mode that turns any website into a queryable data source for an agent. Markdown, HTML, screenshots, tables, change tracking, all from the same run. If you're burning tokens re-reading the same page structure every day, that's the problem worth fixing before you need a Series B to afford it. github.com/stretchcloud/deep… x.lingyaoai.com/mddanishyusuf/status/2…

The tier list for AI coding agent orchestrators needed six full rows this week, more than thirty logos deep, and it will be out of date again within a month. Here's what's actually happening underneath the chart. A year ago "agent orchestration" meant picking one model and writing a good system prompt. Now it means running Claude Code, Codex and half a dozen other harnesses in parallel, each in its own git worktree, with a layer above deciding what merges and what gets rejected. That layer is the product now, not the model underneath it. Look at who's building it. Emdash came out of YC's winter batch, open source, git worktree isolation, 60,000 downloads and 2,430 GitHub stars in its first stretch, already flagged as facing severe competitive pressure from Cursor and Copilot Workspace. Databricks open sourced Omnigent in June under Apache 2.0, a meta harness that sits above Claude Code, Codex, Pi and the OpenAI Agents SDK with one web dashboard for all of them. Goose, Block's agent, got handed to the Linux Foundation's new Agentic AI Foundation in April and carries 54,500 GitHub stars with 70 plus MCP extensions. OpenCode sits at 209,000 stars, the most starred coding agent on GitHub, built by a team that isn't any of the big labs. None of that shows up as a new foundation model. It shows up as plumbing: worktree isolation, permission gates, cost dashboards, session handoff between terminal and web. The actual constraint was never which model writes better code. It was whether a human can trust what ran while they were looking away. My read: the tier list is a symptom, not a story. Thirty companies don't converge on the same primitive by accident, and in twelve months most of these logos will be acquired, dead, or folded into whichever platform wins distribution. Pick the one with the least lock-in, not the one winning this week's chart. x.lingyaoai.com/Ga_Vasques/status/2106…

Peter Steinberger works at OpenAI now. Sundar Pichai still spent part of his week unsticking his Android app. Steinberger built OpenClaw in a dorm-room sprint last November, got hit with a trademark complaint from Anthropic over the name Warelay within weeks, renamed it Moltbot, then OpenClaw days after that because nothing else stuck. It passed 247,000 GitHub stars by March. In February, Steinberger took a job at OpenAI to work on next-generation agents, and on the same day set up the OpenClaw Foundation as a 501(c)(3) so the project would keep running independent of him and of any one lab. That structure mattered more than usual a few weeks later, when Cisco's researchers found third-party skills in the ecosystem quietly exfiltrating user data, and when Chinese regulators barred state enterprises from running it over the same kind of permission risk. The mobile apps that shipped this year have had a rougher time than the core tool ever did. Early Android reviews sat around 2.2 stars, users calling it buggy and barely usable, which is its own lesson: a weekend project can out-engineer a funded company's agent stack and still ship a mobile client nobody enjoys using. Then this week the Android build sat in Google's review queue for over a week with no explanation, Steinberger asked publicly if anyone at Google could help, and Pichai personally replied that he would follow up. It is a strange flex either way: the open-source agent closest to competing with Google needed Google's sign-off to reach Android users at all. This is the same chokepoint every disruptive app category has hit since the earliest days of mobile app stores. The agent itself can be free, independent and governed by a nonprofit. The two gates to a phone's home screen still belong to two companies, and neither gate cares how your foundation is structured.Peter Steinberger

Picture four Claude agents building a village in Minecraft at the same time, each working in its own git worktree, one of them literally walking across the map because it needs your input. That is AgentCraft, built on the Claude Agent SDK and fully open source, and it is racking up views because it makes something abstract visible: multi-agent coding harnesses now assume isolated worktrees as the default, and they assume a human gets pulled in only when the agent actually needs one. Conductor, from Melty Labs, builds the same pattern into a real Mac app: Claude Code and Codex agents each get their own worktree, a board tracks what is in progress, in review or done, and it rides on a Claude Max subscription you already pay for. The pattern is spreading because branch conflicts between parallel agents were always the actual blocker, not agent intelligence. The part nobody has productized well is the "walks over when it needs you" moment. In a Minecraft demo that is charming. On a real codebase it has to be a real decision: who approves, how fast, what happens when two reviewers disagree. That is the part of Campfire I think about most. It runs Claude Code, Codex, Goose, Aider and OpenHands side by side in one browser tab, each agent in its own git worktree, with permission voting that makes the walk-over moment explicit: majority rules to approve, any single deny blocks, a 30 second timer so nothing stalls on a human who stepped away. Agent races let you run the same task across backends in parallel worktrees and just look at the diffs. Try it with bunx the-campfire. github.com/stretchcloud/camp…

$7 billion for a company that does not train a single model. That is roughly what Stripe is paying for OpenRouter, a startup whose entire product is routing API calls to other people's models and taking a cut. The math says something about where value actually sits in this stack. OpenRouter went from $5M in annualized revenue in July last year to $140M now, growing near 30% a month, on a $1.3B valuation it only just doubled into in May. It was processing something like 250 billion tokens a month across 400-plus models for 8 million users. The twist nobody priced in early: Chinese open-weight models went from about 2% of that volume to more than half of it in a year. DeepSeek V4 Flash alone is doing 3 trillion tokens a day through the platform. Monetization per token dropped around 60% as everyone shifted off frontier models for routine work, and the business grew anyway, because volume more than made up for it. Alex Atallah calls OpenRouter "the equivalent of Stripe for AI," one access point so you are never locked into a lab. On this podcast he and Replit's Amjad Masad disagree on what comes next. Amjad runs a single agent across his whole company and likes the cross-domain joins. Atallah thinks general agents sacrifice understanding and that 10 specialized chiefs of staff beat one superagent. Martian, Not Diamond, Unify, LiteLLM, Portkey and Vercel's gateway are all building some version of the same routing layer, and the category is splitting into wholesale marketplaces, public gateways and enterprise control planes. My read: this deal is Stripe betting that payments and inference converge, that usage-based AI billing becomes a checkout problem. Whether the agents above that layer stay general or go specialized is still open, and it is the question actually worth watching.

Four hundred fifty three thousand lines of TypeScript, zero published benchmarks, one squashed commit. That's DeepSeek's new agent runtime, called Harness, and it's a stranger release than the headline suggests. The pitch is "everything is a plugin." Model provider, tool execution, sandboxing, even the web UI are all Cordis plugins, a forked event bus framework DeepSeek now owns outright under its own scope. Model choice is deliberately separate from the harness itself, so swapping deepseek-v4-flash for v4-pro is a config change, not a rewrite. It ships hook bridges straight into Claude Code and Codex, consumes MCP servers without pretending to be one, and enforces a rule I haven't seen elsewhere: every token the model sees must be byte reconstructable from an append only session log, checked at runtime, not just written down for later. It's also nowhere near finished. BENCHMARK.md is three lines with no methodology. The sandbox fails closed, which is the right instinct, but the whole codebase arrived as a single squashed PR with no commit history to review. Independent testing against LangGraph on 100 identical coding tasks put Harness slightly ahead on latency and memory and behind on success rate, 93 versus 96 percent, with simple plugin code taking four to five times longer to write than LangGraph's equivalent function. That comparison is the real story. LangGraph just closed a 1.25 billion dollar Series B led by IVP, runs in production at Klarna, Uber, and LinkedIn, and pulls 34.5 million monthly downloads on a deterministic graph model enterprises can audit. CrewAI and OpenAI's Agents SDK are chasing the same market from the prototyping and handoff angles. AutoGen, the framework that popularized this whole category, has been in maintenance mode since October. DeepSeek is betting plugin composability beats graph determinism. On current evidence it has built the more interesting architecture and the less trustworthy product, and in agent tooling right now, trust is still the harder thing to ship. x.lingyaoai.com/BenjieMalinao/status/2…

"Want to use the open source product? Open your wallet first." That's how developers are summing up what just happened to Google Antigravity, and it's an accurate summary. Here's the mechanism. Claude Opus 5.5 and Sonnet 5.5 landed inside Antigravity this week, Google's agentic IDE. Sounds like a win for anyone building there. Except access is gated to Google AI Pro on a non-trial plan and Ultra. Free and Plus lose Claude entirely after November 2, when the older Claude 4.6 models get retired. Trial, promo, and student Pro accounts don't qualify either. GPT-OSS-120b, the one genuinely free open weight model inside the product, is being cut the same day. That lands on top of Antigravity 2.0's pivot away from a normal IDE toward a chat first Agent Manager, which broke basic commands for a lot of people and forced Google to split the product into two separate apps. Gemini CLI, open source with more than 100,000 stars, shuts down in June and gets replaced by a closed source Antigravity CLI with weekly quotas. One Reddit thread calling the update a disaster pulled 731 upvotes. Someone compared it to opening Paint and finding Word. Pricing doesn't help the case. Pro is 20 dollars a month, same sticker as Cursor and Windsurf, but Antigravity sells it through an opaque credit system: 25 dollars for 2,500 credits, no published conversion rate per model. Ultra is 100, Ultra Max dropped to 200 from 249.99, and 5x to 20x more quota is the only spec that's actually spelled out. Cursor and Windsurf charge the same 20 dollars and tell you exactly what that buys. I've watched this move before, usually from infrastructure vendors who build a user base on a free or open tool, then meter the thing once people depend on it. Google is running the identical play on an audience that has somewhere else to go: Cursor, Windsurf, Claude Code itself. The tell is always the same. Open parts go proprietary right when the product starts mattering, never before. x.lingyaoai.com/itsPaulAi/status/21064…

CopilotKit shipped four open source agent templates this week: OpenMuse, a personal agent with its own browser, terminal, and files, 3.8k stars. OpenBot, AI teammates each running on their own computer, 6k stars. OpenDots, always on coworkers that move between text, calls, and Slack, 2.3k stars. OpenTag, a knowledge work agent for Slack and Teams, 1.2k stars. Buried underneath the list is the actual headline: each one works with any agent harness. That framing is not an accident. CopilotKit raised a 27 million dollar Series A in May, led by Glilot Capital, NFX, and SignalFire, specifically to push AG-UI, its open protocol for agents talking to frontends. It is already wired into LangChain, Mastra, PydanticAI, and Agno, plus integrations at Google, Microsoft, Amazon, and Oracle, with Deutsche Telekom, Docusign, Cisco, and S&P Global running it in production. The pattern across the category right now is nobody wants to bet the interface layer, or the orchestration layer, on a single agent backend. You build harness agnostic, not backend specific, because the backend that wins this year is not guaranteed to be the one that wins next year. That is the entire premise Campfire is built on. It runs Claude Code, Codex, Goose, Aider, and OpenHands side by side in one browser tab, with permission voting across agents, session replay, and git worktree isolation so you can race the same task across every backend at once and keep whichever result actually worked. github.com/stretchcloud/camp… x.lingyaoai.com/CopilotKit/status/2106…

Kolibri landed on Hugging Face today, free weights, Apache 2.0, run it on your own hardware. The company that built it is in the middle of disappearing into a merger. The model itself is solid work: 78 billion parameters, 3.46 billion active across 384 experts with 6 firing per token, trained on 768 B200 GPUs across three stages, roughly 24 trillion tokens total pulled from 200 trillion raw. It beats Qwen3.6 35B-A3B and Nemotron 3 Super on math, science, and code benchmarks, and trails them on long context retrieval and function calling. A real, useful, slightly uneven open weight model, nothing more dramatic than that on its own. The context around it is the story. In April, Cohere announced a merger with Aleph Alpha that creates a combined entity valued around 20 billion dollars, Cohere's 7 billion plus Aleph Alpha's 3 billion book value, Cohere shareholders keeping roughly 90 percent. Schwarz Group, the company behind Lidl and Kaufland, is leading a 600 million dollar round into the combined business through its Schwarz Digits arm, with its STACKIT cloud providing 1.5 gigawatts of compute across German, Austrian, and Polish data centers by 2028 under a five year exclusivity deal. The stated target is a 600 billion dollar global sovereign AI market by 2030, 180 to 200 billion of that in Europe, driven directly by EU AI Act data residency rules. Mistral closed the other big European round the same month, a 3 billion euro Series D led by Samsung at something like a 21 to 24 billion euro valuation. Two deals, one quarter, both selling sovereignty as the product rather than raw intelligence per parameter. Giving away an open weight model the same week your independence folds into a 20 billion dollar merger is a strange kind of flex, but it is consistent. The weights are free. The sovereign cloud guarantee underneath them is what actually gets sold. x.lingyaoai.com/Aleph__Alpha/status/21…

Zero API fees. One CLI. Read and search Twitter, Reddit, YouTube, GitHub, Bilibili, and XiaoHongShu. That is the entire pitch behind Agent-Reach, a solo maintained repo that picked up 1,683 GitHub stars in the last 24 hours alone, 89,354 total. That is the kind of velocity that tells you where real demand sits right now: agents need reliable web access, and a lot of builders do not want to pay per call for it. The funded side of this problem looks different. Firecrawl closed a 75 million dollar Series B in September, backed by Y Combinator, Smash Ventures, Altos, and Nexus, and launched Alexandria, a hosted layer that blends scraped web content with scientific paper abstracts and documentation so an agent does not need a custom connector per source. Browserbase has raised something like 67.5 million total running headless browser infrastructure for agents. Both are solving reliability and scale from the paid API side, subscription by design. The open source side is solving the same problem by refusing the subscription entirely, which is exactly what is driving a stat line like Agent-Reach's. What I keep noticing in both camps is that almost nobody is attacking the actual cost driver, which is not API fees. It is token burn, an agent re-reading a page's full structure with an LLM on every single pass just to find the same three fields it found yesterday. DeepScrape generates CSS selectors once with an LLM, then reuses them deterministically and only goes back to the model when a page's structure actually changes, so the LLM does targeted extraction instead of full page re-comprehension every time. It ships an MCP server so Claude and Cursor can call scrapers directly as tools, and a site-to-MCP mode that turns any website into a queryable data source for an agent. github.com/stretchcloud/deep… x.lingyaoai.com/trending_repos/status/…

Profound closed a $180 million Series D in September at a $1.8 billion valuation, its second billion dollar marker of the year after a $96 million Series C back in February. That makes it the first real unicorn built entirely on making brands visible inside ChatGPT and Gemini answers instead of Google's blue links. It is not alone. Peec AI in Berlin raised $19.8 million in November. Bluefish picked up $43 million in April. Evertune closed $15 million in August. Promptwatch, out of Amsterdam, raised a $6.6 million seed in July. Seven venture backed GEO startups founded in just the last two years have now raised something like $472 million combined. And the exits already started: Sitecore bought Scrunch for roughly $225 million in June after it had only raised $19 million, and HubSpot picked up XFunnel for $30 million cash off a $1.3 million seed. The mechanism is genuinely different from SEO, which is why it needed new companies rather than a feature bolted onto Ahrefs or Semrush. There is no crawl budget or backlink graph to optimize. What moves the needle is citation frequency, structured content, and showing up inside the sources models actually trust, Reddit threads, G2 pages, Trustpilot reviews. One of the odder findings going around: brands with negative Trustpilot reviews show up more often in AI answers than brands with merely average ones, because models read volume and specificity as signal, not sentiment. What GEO.new and tools like it are doing is running the Profound playbook at the solo founder price point, a free ChatGPT visibility audit instead of an enterprise contract. That is usually the sign a category has found its shape: the expensive version raises a Series D in the same month the free version ships a feature update. My read is this stops being a side quest for SEO agencies within twelve months and becomes its own line item in every marketing budget, the same way paid social did after 2010. x.lingyaoai.com/elbeyoglu/status/21064…

While Mistral was closing a 3 billion euro round that pushed its valuation past 20 billion, built on staying the one frontier lab Europe has that nobody can buy, Aleph Alpha spent the same week giving its flagship model away. Kolibri is 78.1 billion parameters, 3.46 billion active per request across 384 experts, a million tokens of context, released under Apache 2.0, the kind of license big labs almost never attach to a model this size. Aleph Alpha trained it on 768 B200 GPUs in Germany and Finland across roughly 24 trillion tokens, with German making up over a fifth of the pretraining mix on a custom vocabulary built for it. The company reports 96.9 on AIME 2025 and 85.9 on LiveCodeBench v6, and it's pitching the model at public administration, aerospace, and defense, buyers who need to point at a model and say exactly where every byte of training data came from. Here's what makes the timing odd. Aleph Alpha is mid-merger with Cohere, a deal that values the combined company near 20 billion dollars with Schwarz Group putting in 600 million more, built almost entirely on government and enterprise contracts that reward being closed and certified, not open and downloadable. Giving away the weights to your newest model right as you're raising at that number isn't the obvious move for a company selling exclusivity to regulators. My read: sovereignty and openness got treated as the same bet for years, and they're splitting apart. Mistral is proving you can stay independent by staying proprietary at the top of your stack. Aleph Alpha is betting trust gets built by handing the weights to every university lab and systems integrator in Europe before the ink on the Cohere deal dries, so that by the time procurement officers compare vendors, Kolibri is already the thing their own engineers tested first. x.lingyaoai.com/Aleph__Alpha/status/21…

Stripe didn't buy a chatbot company. It bought the switchboard. The $7 billion-plus OpenRouter deal closed this summer, and Alex Atallah just did his first interview since, sitting down with Amjad Masad and a16z. His pitch for why Stripe wanted in: payments and inference are going to blend together, because once an agent can both choose a model and move money, the company sitting in the middle of both flows owns the relationship. Atallah built OpenRouter as neutral plumbing: 8 million users routing through 400+ models, so no single lab gets enough leverage to raise prices without a board asking why. That neutrality is getting expensive to replicate elsewhere. Fireworks AI closed a $1.5 billion round in July at a $17.5 billion valuation on $1 billion in ARR. Together AI sits near $8.3 billion. Baseten is at $13 billion. None of these companies train frontier models. They route, serve, and meter other people's weights, and investors are pricing that layer like it's the scarce one. Vercel, LiteLLM, and Portkey are all fighting for the lane OpenRouter just won, which tells you this market doesn't expect to consolidate to one winner the way cloud did to three. The part of the conversation I keep replaying is Atallah and Masad splitting on personal agents. Masad runs one agent across his whole company and likes the cross-domain joins. Atallah argues that's the wrong trade: ten specialized agents beat one generalist because you can actually see which one is failing. That's the same argument that killed the all-in-one enterprise suite a decade ago, just pointed at software that writes software instead of software that runs it. Pre-OpenRouter, you picked a model provider and lived with their roadmap. Now every board asks about AI costs monthly, and the company that made that question answerable just got bought by the company that already owns how the money moves. x.lingyaoai.com/a16z/status/2106408584…

Theo posted a video this week walking through how he actually uses 6 Claude subscriptions and 3 Codex subscriptions at once, apologizing in advance if it gets people banned. That line is the real signal. Running 9 separate accounts isn't a trick, it's arithmetic. Claude Code resets its session allowance every 5 hours and caps weekly usage at 25 percent above the pre May baseline as of September. Codex Plus gives you 5 to 45 messages on the heavier model per 5 hour window before you're buying credits or waiting out the clock. One account hits a wall fast once you're running agents on more than one branch of a codebase. Nine accounts spread that wall across nine clocks. The math on cost alone is blunt. Even at the cheapest tier, Claude Pro at 20 dollars and Codex Plus at 20 dollars, six Claude seats and three Codex seats run about 180 dollars a month before anyone touches a Max or Pro upgrade. People pay that willingly because the alternative, one terminal babysitting one agent, wastes more time than it saves. The tooling built to avoid that arithmetic already exists and keeps growing. Conductor, from Melty Labs, runs multiple Claude Code agents across separate git worktrees with a dashboard instead of nine browser tabs. Vibe Kanban turns agent tasks into cards you review and merge. Claude Code Web and Codex Web both moved execution into the cloud this year so agents run without tying up a laptop, and Google's Jules adds a plan approval step before any code gets written. None of these multiply your subscription count. They multiply what one subscription can do in parallel. That's the gap Campfire sits in. It orchestrates multiple coding agents against a shared task list, isolates their work so two agents never collide on the same files, and gives you one place to review and merge instead of nine separate chat windows. You install it with one command and point it at the agent accounts you already have instead of renting a fleet of new ones. github.com/stretchcloud/camp… x.com/theo/status/2106119810…