Founder & Builder | Building the-campfire.dev, ephemeramail.com, hydranode.ai | AI Agents| 15x Patents | 7x Author | Building for Growth
London, England
- Tweets 14,688
- Following 988
- Followers 3,505
- Likes 3,151
Pinned Tweet
Delete your CLAUDE.md. Not edit it, not trim it, delete the whole file, along with your skills and your hooks, and see what the model does without them. That's Boris Cherny's actual advice from a recent YC appearance, and the Claude Code team has apparently followed it internally enough times that roughly 80% of their own system prompts are gone. His reasoning: most of those rules patched a weakness in whatever model you ran six or twelve months ago. Newer models often don't have that weakness anymore, so the instruction that used to help is now just constraining a more capable model for no reason. His method is almost embarrassingly plain. Delete everything, work normally for a while, watch where the model actually goes off track, add back exactly one instruction when a mistake repeats, and run the whole cycle again at the next major model release. Keep security and compliance rules. Keep the project quirks a model genuinely can't infer from reading the code. Cut folder structures, dependency lists, and architecture summaries, since a capable model can read those off the repo itself faster than it can read your stale description of them.
This is the same problem the rest of the field keeps rediscovering under different names. Anthropic's own long context guidance, OpenAI's notes on attention tradeoffs inside big windows, and Microsoft Research's GraphRAG papers all circle the same fact: a pile of context that only grows, whether it's a system prompt, a vector store, or a knowledge graph, degrades the thing using it, because stale facts sit next to current ones with no signal about which one wins. DeepSeek Harness takes a different structural swing at the same target, building its agent runtime out of swappable plugins so a stale capability gets replaced instead of buried under new instructions.
What strikes me is how little of this is about model capability and how much is housekeeping nobody assigned themselves. Most teams have a deprecation policy for their actual code. Almost nobody has one for the three paragraphs pasted into an agent's instructions eight months ago and never looked at again. How long is your CLAUDE.md, and when did you last delete anything from it?
nitter.siddz.com/suraj_sharma14/status/…
The creator of Claude Code just told everyone to delete their CLAUDE.md.
Not a typo.
Boris Cherny, at YC:
"Every 6 months delete your CLAUDE.md delete your skills, delete your hooks. See what the model does and it might surprise you."
His reasoning:
• Your old rules were fixing weaknesses of older models
• New models don't have those weaknesses anymore
• So your instructions might now be making Claude worse
The method:
1. Delete CLAUDE.md, skills and hooks
2. Work normally and watch where Claude goes off track
3. Add back ONE instruction only when a mistake keeps happening
4. Repeat after every major model release
What to keep:
• Security and compliance rules
• Project quirks Claude can't figure out from the code
What to cut:
• Folder structures, dependency lists, architecture summaries
• Anything Claude can learn by reading your repo
Most people keep adding rules.
The person who built the tool is removing them.
Bookmark this. How long is your CLAUDE.md?
43
85
1,521
293,106
Kill a coding agent mid task and it forgets everything it was doing. That's the dirty secret behind most harnesses shipping this year, and Vercel shipped durable-harness today specifically to make it go away for Pi, OpenCode, Codex, and Claude Code, wrapping each one from the outside so a crash resumes from the last completed step instead of starting over. Model calls and tool calls get logged before they run and read back from that log on recovery, so a finished call never gets billed twice just because the function died underneath it.
This is durable execution finally meeting agent harnesses specifically, instead of agents bolted onto a general workflow engine. Temporal has done durable execution for years through deterministic replay, but agent code usually needs restructuring to fit that model. Inngest took the event driven route, steps and retries wrapped around your functions, pricing free at the bottom and roughly five hundred a month for Temporal Cloud's Business tier versus ninety nine for Inngest Pro. Restate and newer entrants like Kitaru are chasing the same problem from the Python agent runtime side, open source and self-hosted. Vercel's bet is narrower and more useful for this exact moment: skip the rewrite, keep the harness you already use, store state in Postgres, Mongo, Redis, or Vercel's own sandbox.
What nobody is pricing in yet is debugging once agents run for days instead of minutes. A session that resumes three times across two sandboxes and a crashed function needs traces that make that recovery history legible, which is exactly what Vercel pairs this with through Agent Runs. Whoever owns that observability layer ends up owning a lot of the trust.
Coding agents got good at capability fast this year. Surviving a crash without losing a day of context was always the less glamorous half of the job, and it is the half that decides whether anyone actually lets these things run unattended.
nitter.siddz.com/vercel/status/21126538…
Introducing 𝚍𝚞𝚛𝚊𝚋𝚕𝚎-𝚑𝚊𝚛𝚗𝚎𝚜𝚜. Run Pi, OpenCode, Codex, Claude Code, eve, fx, or your own harness durably on Vercel.
With 𝚍𝚞𝚛𝚊𝚋𝚕𝚎-𝚑𝚊𝚛𝚗𝚎𝚜𝚜, agents automatically resume work after failures and run for hours, days, or months. vercel.com/blog/making-every…
3
2
113
Deleting 190 gigabytes of leftover git worktrees shouldn't be a weekly chore, but it is becoming one for anyone running Claude Code's subagent isolation hard. GitHub issues describe the exact same failure mode in detail: one developer hit 615 orphaned worktrees eating 21 gigabytes after a single day of heavy subagent use, then 764 worktrees and 36 gigabytes the next day, leaving under a quarter gig free before a model download. Detached HEAD worktrees never show up in a normal git status, so they pile up silently until the disk fills.
The ecosystem's answer so far is mostly duct tape. People write SessionStart hooks that sweep orphaned worktrees by hand, or reach for Conductor to run several Claude Code agents in parallel on a Mac with its own cleanup layer, or Anthropic's own Agent View dashboard just to watch multiple sessions at once. Claude Flow takes the orchestration route instead, coordinating agents through its own protocol rather than git isolation. None of these treat the worktree itself as a resource with a lifecycle: created, used, and torn down on purpose instead of abandoned.
That's the part I built Campfire around. Forking and branching run through git worktree isolation on purpose, agent races put the same task across isolated worktrees for Claude Code, Codex, Goose, Aider, and OpenHands side by side, and when a race or session ends, the workspace it used ends with it instead of sitting there as an orphan. Permission voting with a thirty second timer stops a runaway agent from doing damage inside that workspace in the first place, and session replay means you don't need to keep the raw worktree around just to remember what happened in it.
Open source, one bunx command to run it. The parallel agent workflow was never the hard part. Cleaning up after it was.
github.com/stretchcloud/camp…
nitter.siddz.com/dennis_kortsch/status/…
1
53
Eighty point seven billion dollars. That is the net loss sitting inside Anthropic's draft S-1 for the first half of 2026, against sixteen point three billion in revenue that grew nearly fifteen times year over year. Read past the headline number and most of that loss, about seventy two billion, comes from a non-cash accounting charge tied to convertible notes, not cash leaving the building. The cash story underneath is different, and more interesting.
Compute and infrastructure spend hit nearly nine billion in six months, more than all of 2025 combined. The filing discloses roughly four hundred fifty billion in cloud commitments stretched over seven to ten years: a hundred eleven billion with Google Cloud, a hundred ten billion with AWS, thirty one billion with Microsoft, plus chip leases through Nscale and Lambda and a TPU arrangement through Broadcom that alone runs past a hundred sixty billion. Count the deals still being negotiated and Anthropic says total spend could reach seven hundred fifty billion. That is a demand bet that makes most cloud backlogs look conservative.
OpenAI is running the same playbook from the other side: reportedly raising thirty billion plus at a valuation north of one point four trillion, specifically because its own IPO got pushed out to 2027. Two labs, same capital intensity, same hyperscaler dependency, different timing on when public markets get to see the math.
The filing also says plainly that a catastrophic AI-related incident could produce liability exceeding available insurance. That line sits next to customer concentration risk, twenty percent of first half revenue from two customers, and language about hyperscalers increasingly acting like competitors rather than landlords. The revenue growth is real. So is the fact that this business runs on compute contracts nobody can cancel.
My read: at this scale, an IPO stops being a funding event and starts being a disclosure event. Model labs now carry balance sheet risk that looks less like a software company shipping code and more like a utility financing a power plant years before the demand shows up to pay for it.
nitter.siddz.com/WSJ/status/21126281418…
Exclusive: Anthropic's draft IPO filing reveals red-hot growth, massive computing costs and an $80 billion net loss. Here are the highlights. on.wsj.com/3WmyWAR
2
77
Hey @grok should a Vibe coder really know the difference between putting a / at the end or not?
1
33
Hey @grok when do you think Anthropic and OpenAI can solve world hunger? Asking for poor children!
1
2
81
Drug development takes 10 to 15 years from target to approval, and today OpenAI decided the fix isn't a smarter model, it's project management. The company announced it has acquired Orchestra Bio, a platform where biopharma teams and AI agents coordinate the entire pipeline together: when one study slips, the agents trace which other trials and deadlines it touches, gather the relevant data, and hand scientists a set of real options instead of a status update. It folds straight into Rosalind, the life sciences stack OpenAI has been quietly building since last year.
This is the third serious AI-for-drug-discovery bet reshaping the category in under eighteen months, and the price tags show how differently each lab is playing it. Alphabet's Isomorphic Labs, the DeepMind spinout built on AlphaFold, raised 2.1 billion dollars and is reportedly back in talks for a round that would value it at 40 to 50 billion. Recursion Pharmaceuticals went the consolidation route instead, folding Exscientia into itself in an all-stock deal worth 688 million after both had a rough run as public companies on their own. OpenAI's move looks smaller than either, an acquihire wrapped around a coordination layer rather than a discovery engine.
Anthropic has started making similar noises about science and healthcare without a deal to point to yet. What stands out to me is the size gap: Isomorphic is playing for a discovery moat worth tens of billions, Recursion played defense through consolidation, and OpenAI bought the least glamorous layer in the stack, the one that manages the hundred adjacent decisions a trial generates every time something changes. The hard part of drug development was never only the chemistry. It was coordinating fifty people across a shifting timeline, and that happens to be a problem language models are unreasonably good at.
nitter.siddz.com/OpenAI/status/21126073…
We’ve acquired Orchestra Bio, a platform that helps biopharma teams coordinate the complex process of drug development.
Together, we’ll build on Rosalind, our AI models and tools for life sciences research, to streamline those operations and help new treatments reach patients sooner.
Welcome @ashokaraj_ and team!
openai.com/index/openai-to-a…
2
1
122
Every hyperscaler racing to dent Nvidia's lead just converged on the same move within a day of each other. Google open sourced TorchTPU at the PyTorch Conference in San Jose this week, co-built with Meta's PyTorch team, letting existing PyTorch code target TPUs through torch.compile with what Google says are minimal changes. A day earlier Amazon shipped TorchNeuron in private beta, same pitch, aimed at Trainium: run your PyTorch code unchanged, no rewrite into JAX or Neuron's own primitives required.
The reasoning is identical on both sides. Something like 85 percent of AI researchers write in PyTorch first, and every chip that isn't CUDA has spent years asking developers to learn a second toolchain just to use it. That tax is what actually slows adoption, not raw FLOPs. TorchTPU already has real numbers behind it: 52 percent model FLOPs utilization on Qwen3 8B, scaling near linearly across 8,192 Ironwood TPUs, Hugging Face's Transformers v5 shipping device_map="tpu", and vLLM, SGLang and Ray already wired in.
Here's the part the announcements skip. AWS claims Trainium beats comparable Nvidia instances by 30 to 40 percent on price performance, but independent benchmarking this year put the real gap closer to a third of that once someone actually measured it instead of reading the spec sheet. Google's TPU v6e only beats a B200 on bandwidth cost if you commit for three years; on demand, it's pricier. AMD's MI355X undercuts B200 by as much as 27 percent on some workloads and barely at all on others, depending entirely on which inference framework you pick.
None of that will slow the marketing. What it tells me is the real competition quietly moved up a layer, off the chip spec sheet and onto whoever makes their silicon forgettable to the person writing the training loop. Nvidia's moat was never the hardware. It was that nobody wanted to learn XLA.
nitter.siddz.com/MainzOnX/status/211261…
1
114
Half of this is already true, and it’s the half that cuts against the rest.
Local specialised coding models are here. JetBrains shipped Mellum 2.1 this month: 12B parameters, 2.5B active, Apache 2.0, runs on a laptop through Ollama, 47 on SWE-bench Verified. Domain players are already priced like winners too, Harvey at $11 billion in legal, Harmonic at $1.45 billion in formally verified math.
But look at who actually distributes AI to developers. GitHub Copilot has 50 million users, and since April it runs Anthropic’s Claude agent and OpenAI’s Codex agent directly inside github.com, with Opus and GPT Codex models included in the subscription. Microsoft’s biggest developer channel is reselling the labs, not routing around them.
The demand side doesn’t look like a company about to fail either. Anthropic told investors its run rate hit $65 billion in July, up from roughly $9 billion at the end of 2025. OpenAI is reportedly past $40 billion annualized.
Where I’d take the bubble argument seriously is the balance sheet. OpenAI raised $122 billion at an $852 billion valuation in March and has reportedly committed around $600 billion to compute through 2030, much of it take-or-pay, while burning upwards of $15 billion a year. If something breaks in the next three years, my bet is it breaks there, in contracts signed against revenue that hasn’t arrived yet, while usage keeps climbing every quarter.
My AI predictions.
In the next 1 to 3 years, Open AI or Anthropic, or both, fail / get acquired for a fraction of their current valuations.
Microsoft, Google or a developer tooling company with access to lots of developers and source code like Sonar or JetBrains wins AI in software engineering.
Probably with smaller, local, specialised models.
Other domain specific companies spring up to win domains such as mathematics, legal, writing, etc.
TL/DR: The technology isn’t going away, but the financial bubble will burst.
What do you think?
1
1
118
Stripe is now being described as hard to use by an unlikely group: AI agents. Lightsage, a two-person company that just raised 4 million dollars led by Nexus Venture Partners, runs a public benchmark called DevToolArena that scores how well a coding agent can discover, sign up for, and integrate a developer tool entirely on its own, no human in the loop. The categories read like a normal vendor landscape: auth, voice, inference, storage, search, payments. The scoring doesn't.
One builder ran Cursor's Composer 2.5 through it across 120-plus live production APIs, each with its own isolated sandbox and a real integration task. Median time to a working integration: 1.3 minutes. Codex came in at 1.8, Gemini at 9.6. A five-way tie for the top category score went to Deepgram, ElevenLabs, Fireworks AI, Prefect and Reducto. Cursor also needed the fewest web searches per task, about 3, against 7.7 for Grok's agent and 5.6 for Codex, which says as much about how well those APIs document themselves for a machine reader as it does about the agent doing the reading.
A separate project, the AX Benchmark, scores email platforms like Resend, SendGrid and Postmark the same way: discoverability, sign-up accessibility, onboarding clarity, API design, SDK quality, each on a 1 to 4 scale, using Claude Code to attempt a real integration from nothing.
What both are measuring is a genuinely new failure mode. A tool can have perfect documentation for a person and still be nearly invisible to an agent if the signup flow needs a CAPTCHA, the API keys live behind an email confirmation, or the SDK examples assume a browser session the agent doesn't have. Developer relations used to mean writing a good quickstart. Now there's a second audience reading that quickstart who can't click a verification link, and most vendors haven't noticed they have a second audience yet.
nitter.siddz.com/jun_liang_sf/status/21…
Can Cursor actually use your API?
I ran Cursor's agent (Composer 2.5) through DevToolArena. 120+ live production APIs, a real-world developer task for each, and its own isolated sandbox.
It's also the fastest: a median of 1.3 minutes per task. Codex, the next fastest, takes 1.8. Gemini takes 9.6.
Top 🏆: a five-way tie at 97: Deepgram, ElevenLabs, Fireworks AI, Prefect, Reducto
Category leaders on Cursor:
- Voice STT: @Deepgram
- Voice TTS: @ElevenLabs
- Inference: @Fireworks
- Durable Workflow: @PrefectIO
- Document Parsing: @reductoai
- Auth: @clerk and @WorkOS
- Sandboxes: @daytonaio and Vercel
- Storage: @Box
- Vector DB: @qdrant_engine
- Graph DB: @memgraphdb
- Payment: @stripe
- Search: @firecrawl
- Browser: @Steel
- Voice Telephony: @bolna_dev
- Voice Infra: @livekit
What stood out
- Cursor searches less: about 3 web searches per task, against 7.7 for Grok and 5.6 for Codex.
Full evals at DevToolArena.com/cursor. Scores are based on discoverability, usability, tool calls, and errors.
Comment "Agent" and I'll send you your report. Repost and tag your favorite API so I can add it to the list.
*All runs are independent and in isolated sandboxes, powered by @lightsagehq AX infra.
6
1
3
181
Eleven and a half points. That's what one round of self-editing bought GLM-5 on a legal reasoning benchmark, with zero retraining, using a tool Google just open sourced called Harness2.
Harness2 lets an agent rewrite its own scaffolding: the system prompt, memory, tool guidance, subagents, scripts, on every task. It proposes a change to the harness in front of it, tests the change on that same task, keeps whatever actually worked with no answer key needed, and carries the lesson into future tasks. Apache-2.0, already past 11,000 stars, works with OpenCode or Codex out of the box, and in Google's own tests it picked up 11.5 points on WorkBuddy-Bench and 11.8 on LAB using GLM-5 and Gemini 3.5 Flash.
This isn't an isolated idea, it's a pattern forming across several labs at once. An academic project called Self-Harness runs a near-identical loop on Terminal-Bench-2.0 and took MiniMax M2.5 from a 40.5% to a 61.9% pass rate without touching a single weight. Researchers at Xiaomi built HarnessX, which treats the harness as swappable modules run by four cooperating agents, a Digester, Planner, Evolver and Critic, and reports a 14.5% average gain across ALFWorld, GAIA and SWE-bench Verified from harness evolution alone, plus another 4.7% once they let the model co-evolve too. Google DeepMind has a parallel paper, Evo2, that goes further still and interleaves harness edits with actual model training, landing a 48.3% resolved rate on SWE-bench Verified.
Here's what I keep coming back to: Cursor, Aider, Cline and Claude Code are all hand-tuned harnesses today, built once by a person and then frozen until the next release cycle. What Harness2 and its cousins are proposing is that the harness stops being a one-time design decision and becomes a training signal the agent maintains on its own, the way a team maintains its own runbook.
There's a ceiling on both sides worth naming. Tune only the harness and you cap out at whatever reasoning the base model can already do. Train only the model and you cap out at whatever the harness bothers to ask it for. The real work right now is happening in the seam between those two, and that's a genuinely different skill than prompt engineering was a year ago.
nitter.siddz.com/RoundtableSpace/status…
Google open-sourced Harness². It lets an AI agent improve its own setup on every new task, without retraining the model.
The setup around a model, its "harness", is normally designed once by a person: the system prompt, guardrails, memory, tool guidance, subagents, skills, scripts and plugins. Harness² edits all 8 while it works.
How it works:
→ proposes changes to the harness for the task in front of it
→ tests those changes on that same task
→ keeps what actually worked better, with no answer key needed
→ carries the lessons into future tasks
In Google's tests, one round improved GLM-5 and Gemini 3.5 Flash by up to 11.5 points on WorkBuddy-Bench and 11.8 points on LAB, a legal benchmark.
Works with OpenCode or Codex. The demo runs on Linux with no API keys:
git clone github.com/google-research/h… && cd harness2
uv venv --python 3.12 .venv && source .venv/bin/activate
uv pip install -e '.[render]'
harness2-demo --output runs/demo --k 2
Apache-2.0, 11.3K stars.
4
307
Lead your agents like an orchestra, one post on building AI agents argued this week, alongside five other good principles about giving agents separate workspaces and keeping humans in charge of what ships. Nice framing. Most of the tools people actually reach for to do that can't get two agents to acknowledge the other one exists.
I went through the current crop of open source coding agent orchestrators recently. Agent Orchestrator runs 26 harnesses in separate git worktrees, but agents sit idle waiting on your approval. Emdash spins up 34 CLI providers in parallel and says plainly that agents don't coordinate, you review and merge yourself. Claude Squad combines tmux with worktrees for Claude Code, Codex, Aider and Gemini side by side, and its autoyes flag that skips approval prompts is still labeled experimental, which is exactly the flag you'd want not to be experimental. Crystal, one of the better known tools in this space, got deprecated in February. Vibe Kanban lost its original backer and is now community maintained.
The pattern across all of them: isolation solved, coordination mostly ignored, and permission handling stuck at either babysit every prompt or turn approval off entirely.
That's the gap Campfire runs at directly. Permission voting across agents, majority rules approves, any single deny blocks, a 30 second timer so nothing hangs waiting on a human who stepped away. Agent races, the same task run across Claude Code, Codex, Goose, Aider and OpenHands in isolated worktrees so you can compare outputs instead of picking a harness on faith. A Collective Intelligence layer gives agents shared context and semantic memory instead of five sessions that never learn what the others already tried. One bunx the-campfire install, one browser tab, every backend you're already paying for.
github.com/stretchcloud/camp…
nitter.siddz.com/MaryamMiradi/status/21…
🛠️🧭 "Slop-Free Zones for Building AI Agents" got me digging.
If you are building with AI coding agents like Claude Code or Codex, you have probably struggled with:
Messy codebases, endless workflow tweaking, lost context, risky AI-generated changes, and multiple agents that are difficult to coordinate.
I watched an AI Engineer talk about Slop-Free Zones, where human review is non-negotiable.
So I pulled together the 6 principles from the talk and matched each with a practical GitHub resource.
Here are 6 principles & GitHub repos to build AI Agents faster:
》𝗣𝗿𝗶𝗻𝗰𝗶𝗽𝗹𝗲 𝟭: Keep Trying New AI Models
✸ Try new AI models and coding tools when they come out.
✸ Compare which ones work faster, cost less, and make fewer mistakes.
✸ Spend more time building than testing tools.
→ Dont waste time building workarounds that AI tools will likely offer as standard features.
》𝗣𝗿𝗶𝗻𝗰𝗶𝗽𝗹𝗲 𝟮: Don't Overcomplicate Your Setup
✸ Use existing tools instead of building everything yourself.
✸ Create custom workflows only when your project needs them.
✸ Avoid spending hours tweaking tools instead of writing code.
→ Only build custom tools when your project needs something existing tools cannot do.
》𝗣𝗿𝗶𝗻𝗰𝗶𝗽𝗹𝗲 𝟯: Create Slop-Free Zones
✸ Decide which files AI agents should never change without approval.
✸ Require human review for database changes, security settings, and important code.
✸ Keep AGENTS. md, CLAUDE. md, and skill files clear and accurate.
→ One bad instruction can cause your AI agent to make the same mistake again and again.
》𝗣𝗿𝗶𝗻𝗰𝗶𝗽𝗹𝗲 𝟰: Give Agents the Right Information
✸ Collect useful information from Slack, meetings, documents, and bug reports.
✸ Store it where agents can search and find it.
✸ Let agents look up what they need instead of giving them everything at once.
→ Let agents search for the information they need instead of putting everything into their prompts.
》𝗣𝗿𝗶𝗻𝗰𝗶𝗽𝗹𝗲 𝟱: Let Agents Work Independently
✸ Run agents in safe, separate workspaces.
✸ Let them keep working even when your laptop is closed.
✸ Allow agents and people to work together without getting in each other's way.
→ Give agents safe places to work so they can keep going even when your laptop is closed.
》𝗣𝗿𝗶𝗻𝗰𝗶𝗽𝗹𝗲 𝟲: Lead Your Agents Like an Orchestra
✸ Give different agents different tasks.
✸ Check their work before combining changes.
✸ Keep people in charge of important decisions.
→ Let agents do the work, but keep people in charge of what gets built and shipped.
So, next time you build AI Agents:
Try New Models →
Keep It Simple →
Slop-Free Zones →
Right Information →
Let Agents Work →
Orchestra Mode
𝗣𝗹𝗲𝗮𝘀𝗲 𝗥𝗲𝗺𝗲𝗺𝗯𝗲𝗿:
The best builders give agents freedom where experimentation is cheap and enforce human review where mistakes are expensive.
ALT Colorful infographic titled "Build AI Agents in Slop-Free Zones" by Dr. Maryam Miradi, featuring six principles for building reliable AI agents: Try New Models, Keep It Simple, Slop-Free Zones, Right Information, Let Agents Work, and Orchestra Mode. Each principle includes a practical explanation and a relevant GitHub repository for AI engineers. The visual uses bold hand-drawn typography, colorful labels, and a playful architectural background, with a reminder that humans stay in charge of important decisions.
1
2
149
Seventy percent of Tolan's users are women. Forty percent are parents. One in six is caring for an aging parent. That is not the demographic chart anyone drew up when AI companion apps got written off as a lonely young men category, and it's part of why Union Square Ventures just led a 30 million dollar round into Portola, the company behind it, alongside Khosla Ventures and Homebrew.
The round roughly doubles Tolan's total funding to about 60 million, on top of the 10 million seed and 20 million Series A it had already closed. PitchBook puts the new valuation near 268 million, up from 120 million. Annual recurring revenue has more than doubled year over year into double digit millions, monthly active users sit around 250,000, and the company logged more than 70 million conversations over the past year. Thirty day retention improved 115 percent year over year, the number that actually explains why investors who backed the Series A came back for more.
Tolan isn't the biggest name in this category. Character.AI still dwarfs it on raw usage, Replika has years of head start, and Meta shipped its own Muse companion into the same market this year with a trillion dollar balance sheet behind it. What Tolan is selling instead is retention data and a user base nobody else is courting as directly, parents and caregivers rather than teenagers killing time between classes. Quinten Farmer sold his last company, a fintech called Even, to Walmart for 300 million, which probably explains why this pitch reads like a retention business rather than a chatbot demo.
The 94 percent of surveyed users who said they turn to Tolan when people in their life are unavailable cuts both ways. Investors read it as proof of a daily habit worth funding. A psychologist reads it as a signal the product sits closer to a crutch than a bridge. Tolan has clearly read both memos, since it just added crisis safeguards and keeps nudging users toward real relationships inside its own app. That tension, not model quality, is the actual product risk here.
1
53
Ten repos deep into a "tools that make web scraping powerful" thread this week: Crawlee, Scrapy, Playwright, Colly, trafilatura, Beautiful Soup, even an Oxylabs MCP server for bringing extraction into agent workflows. Good list. Every one of them solves fetching and parsing. None of them solve what happens the day a site's HTML changes and your selectors silently start returning garbage.
That gap is why Firecrawl just raised a 75 million dollar Series B led by Smash Capital to build Alexandria, a knowledge library pulling structured data from something like 70 million repos and tens of millions of government filings. Bright Data leans on 400 million residential IPs. Apify ships 49,000 prebuilt actors on top of open source Crawlee. Diffbot does knowledge graph extraction starting at 299 dollars a month. Everyone is racing to own the raw collection layer, and almost nobody is racing to make that data something an agent can query without burning tokens re-reading the same page on every call.
That's the part I built DeepScrape around. Selectors get generated once with an LLM, then they self-heal when a page's structure shifts instead of triggering a fresh model read every time. The LLM layer only gets called for the targeted extraction and validation work it's actually good at. Site to MCP turns any site into a structured source Claude or Cursor can query directly, instead of another tool call that dumps back a wall of markdown.
Open source, Playwright underneath, built for agents that need web data to be reliable and cheap, not just fetched.
github.com/stretchcloud/deep…
nitter.siddz.com/DivyanshT91162/status/…
10 GITHUB REPOS THAT MAKE WEB SCRAPING SERIOUSLY POWERFUL
The web is full of valuable data.
The challenge is collecting it efficiently, structuring it properly, and turning it into something useful.
Here are 10 repos worth exploring:
1. Crawlee — Build reliable crawlers with JavaScript or TypeScript.
github.com/apify/crawlee
2. Scrapy — A battle-tested framework for large-scale crawling.
github.com/scrapy/scrapy
3. Playwright — Control browsers and automate complex web interactions.
github.com/microsoft/playwri…
4. Oxylabs Web API MCP — Bring web data extraction into MCP workflows.
github.com/oxylabs/web-api-m…
5. Colly — A fast web scraping framework for Go.
github.com/gocolly/colly
6. Newspaper4k — Extract and parse news articles from websites.
github.com/AndyTheFactory/ne…
7. trafilatura — Extract main text and metadata from web pages.
github.com/adbar/trafilatura
8. Beautiful Soup — Parse HTML and XML without the headache.
github.com/wention/Beautiful…
9. MechanicalSoup — Automate website interactions with Python.
github.com/MechanicalSoup/Me…
10. Browserless — Run browser automation and scraping in the cloud.
github.com/browserless/brows…
The best data pipeline starts with the right tools. Save this list.
3
72
Scroll past rank ten on the image to video leaderboard and you start finding the names people actually recognize. Veo 3.1 sits at fourteen. Kling 3.0 Pro is nineteen, Elo 1000, literally the anchor point the whole scale is pinned to. The top of the list belongs to models most casual users have never typed into a search bar.
FLUX 3 Max just landed in second place on Artificial Analysis's video arena, Elo 1138, two points behind MiniMax H3 and tied with ByteDance's Dreamina Seedance 2.5. Built by fal Research with Black Forest Labs, it turns a 10 second 720p clip around in about 12 seconds of generation time, and fal priced it at 75 percent off for three weeks, the cheapest model anywhere near that quality tier right now.
What that leaderboard is quietly saying: the video model race stopped being a two company story a while back. Google and OpenAI get the headlines, but Gemini Omni Flash, Alibaba's Wan 3.0 and MiniMax's H3 line are the ones actually trading places at the top. Sora and Runway don't even show up on this particular board, which is its own data point about where third party benchmarking attention has drifted.
The gap between brand recognition and leaderboard position is becoming the real story in generative video. Pricing is doing a lot of the sorting too. A 75 percent discount on a frontier tier model is a land grab, not a victory lap, and whoever gets developers building against their API before that window closes wins the next six months here, not whoever wins the next headline.
1
1
162
Model weights are free. Running them at scale cheaply is not, and that's the part DeepSeek open sourced for everyone this week.
DeepInfer is the inference engine serving DeepSeek V4.1 Flash in production right now, and it's fully open, alongside deepseek-recipe, the token encoding component underneath it. The feature list reads like a serving team's wishlist: distributed inference, speculative decoding, structured outputs, and something called SWA Bounded Replay for sliding window attention. None of that sounds exciting until you remember model weights alone don't make AI cheap, the serving stack does, and most labs treat that stack as the actual secret.
The detail I keep coming back to is hardware. DeepInfer runs on Nvidia and on Huawei Ascend, and DeepSeek has reportedly committed to at least 160,000 Ascend 950DT chips for a data center in Inner Mongolia. Huawei is pricing those chips 60 percent higher on supply constraints and still can't make enough. Pair that with open serving code tuned for Ascend and you get a credible non Nvidia inference path for the first time, not just a chip announcement.
This drops into an inference engine market that already has vLLM as the default for high concurrency serving, SGLang gaining ground on agent pipelines with heavy prompt reuse, and TensorRT LLM locked to Nvidia hardware. DeepInfer doesn't claim to beat any of them on a chart. It claims to be what's actually running a frontier scale model today, for free, on hardware the others don't fully support.
The open question nobody's answering yet: whether Ascend inference holds up at the throughput DeepSeek is claiming once people outside the company start running it themselves.
nitter.siddz.com/heyshrutimishra/status…
WTF DeepSeek just gave away its production inference stack.
Big move.
DeepInfer, the engine behind DeepSeek-V4.1-Flash, is now open source, plus a component that makes community deployments match the official API.
DeepSeek’s stated goal is lower compute costs and cheaper AI for everyone.
1
194
"Any proper solutions?" is the exact question at the bottom of a tweet this morning about agents re-learning an entire project from scratch every new session. The complaint underneath it is the real one: AGENTS.md goes stale the moment nobody updates it, and even when it's current, re-reading it still burns tokens every single run.
I've watched this exact failure mode take down more agent workflows than any model limitation has. The fix isn't a better markdown file, it's treating what an agent learned about a codebase as state that persists and gets shared, not a document that gets re-read.
That's the layer Campfire runs underneath Claude Code, Codex, Goose, Aider and OpenHands when you run them side by side in one tab. Semantic memory and shared context streams carry forward what an agent already figured out about a repo, so the next session, and the next agent, isn't starting from a blank page. Session replay means you can see exactly what state carried over and why, instead of trusting a stale file.
Install is bunx the-campfire. No AGENTS.md archaeology required.
github.com/stretchcloud/camp…
nitter.siddz.com/serious_mehta/status/2…
Every time I start a new session, my agent has to re-learn what everything in my project it from scratch. This is such a repetitive waste of tokens. AGENTS [dot] md doesn't solve this either because it's hard to keep it up to date and still burns tokens.
Any proper solutions?
1
1
1
188
Skyfall's launch thread says SKY-001 beats every frontier model tested on CAD and BIM. Skyfall's own research post is more careful: it "matches or exceeds" them. That gap between the tweet and the paper is the whole story here.
The real numbers: on CAD, SKY-001 scores 79.7 percent mean IoU against GPT-6 Astra's 79.1, a 0.6 point edge, while running in 3.9 seconds at $0.005 a case versus 199 seconds and $0.75. On BIM it's 43.6 versus 40.6, with cost dropping from $0.446 to $0.0024. The speed and cost gaps are the believable part. A sub half overlap mean score on BIM means the best output still misses the reference more often than not.
This fits a pattern I'm watching across the category: purpose built models undercutting generalist frontier models on narrow, well defined tasks. Same logic behind smaller coding models beating GPT class systems on single languages, or fine tuned retrieval models beating general embeddings. Skyfall is the team behind Maluuba, the Montreal lab Microsoft bought in 2017, backed by Inovia Capital and Garage Capital. Their other current bet is buying a small business and installing an AI CEO to run it, so the engineering model is one of two wildly different "world model" experiments running at once.
What's missing: the benchmark is self designed, tasks and prompts chosen by Skyfall, judged partly by an LLM panel that includes a competitor model. GPT-6 Astra returned no usable output on 18 percent of assets, which skews the comparison further. Access is waitlist only, so nobody outside the company has run this themselves.
150x cheaper is worth taking seriously if someone outside Skyfall reproduces it. A 0.6 point accuracy edge marketed as "beats every frontier model" is a press release wearing a benchmark's clothes.
nitter.siddz.com/spisallyouneed/status/…
Introducing SKY-001, the first Engineering World Model.
On CAD and BIM tasks, it beats GPT-6 Astra and every frontier model we tested:
-> 51-125x faster
- > 150-210x cheaper
-> Higher IOU than all frontier models
SKY-001 lets you dream a physical object before you build it.
18 months of work. Built by the team out of Maluuba's deep learning lab (acq. by @MSRResearch).
The first ‘AI General Engineer’ will be the backbone for Autonomous Manufacturing.
1
174
GPT-6.1 Sol goes from 5.4 percent accuracy on COBOL migration to 86.6 percent, in the same model family, just by moving up a generation. Push the reasoning effort higher and it hits 97.9, which OpenAI's own writeup calls close to saturating the evaluation entirely.
That's the headline from OpenAI's new push into COBOL, Fortran, Visual Basic and OCaml. The mechanism is MultiLangBench, a 20 language SQLite rebuild test. GPT-6 Astra averages 87.4 percent across all twenty, up from 26.8 for the prior generation. COBOL is still the hardest language in the set at 72 percent, but that's up from 10.9. The cheap option, GPT-6 Luna, runs about 78 percent cheaper per sample than the older model while beating it everywhere.
Here's the part that actually moved markets. Anthropic published a COBOL discovery and risk assessment workflow inside Claude Code back in February. IBM has Project Bob, a VS Code fork wrapping Claude, Mistral, Llama and its own models with FedRAMP and HIPAA context built in, arguing modernization is a platform problem, not a language problem. When that comparison started circulating, IBM shares fell 13 percent in a day and 27 percent for the month, dragging Accenture and Cognizant down with them. Three vendors, one forty year old pile of business logic, and the market just priced in who it thinks loses the integrator relationship.
IBM's own numbers show how unglamorous the real work still is: one customer, NOSI, cut COBOL analysis time by 94 percent using watsonx, and that's just understanding the old code, not migrating a line of it yet.
None of the three has shown byte for byte behavioral equivalence between legacy and migrated systems, which is the actual bar for production mainframe work. Benchmarks measure test pass rate. Nobody's measuring the thing banks actually need proven.
nitter.siddz.com/OpenAI/status/21123103…
Our latest model advances bring stronger capabilities for building and migrating software in low-resource programming languages such as COBOL, Fortran, and Visual Basic.
These languages power software that banks, governments, and businesses depend on. Progress here means understanding decades of business logic and helping developers modernize those systems while preserving what works.
We’re building these capabilities around the challenges developers and enterprise customers bring us. If your team is working on these systems, tell us which languages and workflows we should tackle next.
openai.com/index/low-resourc…
1
82
Every leading image, video and audio model now shares one address. Sunra launched an MCP server this week that puts Veo, Kling, MiniMax, Seedance and the rest inside Claude, ChatGPT or whatever assistant you already run, billed at each maker's own API price with no markup layered on top.
The pricing spread those models hide is the real story. Veo 3.1 runs $0.40 per second at standard quality, $0.60 at 4K, with a Fast tier down at $0.10 to $0.12. Kling 3.0 Turbo prices 720p at roughly $0.11 per second and pushes 1080p into a separate credit plan. MiniMax's newest model lands at $0.13 per second at 2K. Sora 2's API prices $0.10 to $0.70 per second by tier and sunsets entirely on September 24. None of these numbers live anywhere a builder can compare side by side until someone puts a router in front of all of them.
That router is a familiar play. OpenRouter ran this exact pattern for text models and it worked: a $113 million Series B in a year that took it from roughly $19 million to $50 million in annualized revenue, now routing more than 3 trillion tokens a day for 8 million developers across 400 plus models, priced at a $1.3 billion valuation. It's the same move cloud infrastructure made a decade ago, when nobody wanted to integrate with a dozen storage and compute vendors directly, so a routing and gateway layer absorbed that pain and quietly became the default integration point. MCP is trying to replay that for every model on the market at once, text included.
I don't think a no markup router stays no markup for long. OpenRouter's margin comes from volume and provider deals it doesn't advertise. A new entrant promising zero markup on day one is either subsidizing distribution to win the integration point first, or betting the real money shows up later, in enterprise contracts and whichever labs want the traffic badly enough to pay for placement. The model made the news. The router is where the business actually gets decided.
nitter.siddz.com/sunra_ai/status/211221…
Today we're launching Sunra MCP, with the most accessible pricing in the industry.
Every leading image, video and audio model, right inside Claude, ChatGPT and the AI assistant you already use. Each one at its maker's own API price, with no markup added.
Now is the time to build. Go create something no one has seen before.
1
157