Claude Code Writes 90% of Itself as DeepSeek Ties Opus on Browser Tasks for 19x Less
@ClaudeDevs says Claude now writes about 90% of Claude Code's own code, with merged PRs doubling while reverts fell, and @addyosmani argues verification has become the real bottleneck. @browser_use benchmarked DeepSeek V4.1 Flash 1017 level with Opus 5.5 on long-horizon browser use at a 19x lower cost per task, while a live Plex exploit chain and layer-collapse research rounded out the day.
Quick Hits
- Claude writes roughly 90% of Claude Code's code, per @ClaudeDevs: from June to September, merged PRs more than doubled while the reverted share fell over 40%. @addyosmani's read: when agents generate code faster than you can review it, "the bottleneck shifts to proving that it works."
- @browser_use reports DeepSeek V4.1 Flash 1017 matched Opus 5.5 on long-horizon browser use (75/100) at $0.22 per task versus $2.35; @0xSero says he has already swapped GPT for DeepSeek on browser automation.
- If you run Plex at home, update today. @c7five says exploit code for a four-bug chain giving unauthenticated code execution went public two days ago, and attackers dropped a Monero miner on a honeypot within a day. Patched builds: 1.43.3.10861 or newer.
- @MercityAI's LaCo (Layer Collapse) removed 6 layers from Qwen3-8B-Base for a 16.7% smaller model that kept 78-94% of its reasoning scores; @arnavmehta007 speculates the technique, plus quantization and ternary ops, could enable much smaller on-device models.
- @alexocheema argues M5 Ultra's memory economics ($4.5 per GB/s) beat DGX Spark by 5.6x and DGX Station by 4.7x, with a 512GB Mac Studio estimated from $15,899; @twid's takeaway is to preorder before prices move.
Agents write the code; humans prove it works
The day's strongest thread: authorship is getting cheap, proof is not.
@addyosmani, quoting @ClaudeDevs, says Claude Code wrote 90%+ of its code once a verification process was in place, crediting @bcherny and team. One human reading line-by-line no longer works as a gate, he argues; review has to become a multi-layered verification system with fast deterministic feedback: e2e tests, feature maps (as @poteto advocates), property tests, and multi-agent adversarial reviews. The human role becomes product intent, risk assessment, and evaluating evidence, what he calls "judgement, direction, and ownership."
@ibuildthecloud predicts that within a year "anyone can build a software factory like this with some markdown and basically a harness," claiming it is 10x easier than people think. The referent is @BHolmesDev's writeup of a multi-agent factory: a triage agent, an implementation agent, a verification subagent, a code review agent cycling before human review, and monitoring automation that files issues from alerts. @BHolmesDev is now benchmarking whether separate agents are even needed, or whether one high-intelligence model can handle most steps.
For people repositioning, @yacineMTB points to @Tim_Dettmers' blog (via @code_star) on navigating academia and industry in the age of agents, citing Joseph Suarez as proof you can "just do frontier work" outside institutions. @dillon_mulroy recommends @just_be_dev's piece on code review. @kitlangton, meanwhile, teased what he calls his biggest productivity gain since speech-to-text, now open source with Linux and Windows support coming.
Token thrift: delta snapshots, collapsed layers, $0.22 browser tasks
Tooling, benchmarks, and research posts all attacked the same cost line.
@ctatedev added snapshot --delta to agent-browser, returning only what changed since the last snapshot instead of the whole page, which cut tokens 73x when filling a form. @0xSero's switch to DeepSeek ("8x faster, 10x cheaper" by his account) is the user-side echo of @browser_use's benchmark tying DeepSeek V4.1 Flash 1017 to Opus 5.5 at 75/100, with Sol still SOTA but the gap closing.
On the research side, @MercityAI says LaCo removed up to 10 whole layers from Qwen3-8B-Base; deleting 6 made the model 16.7% smaller while keeping 78-94% of reasoning scores at a 1.68x perplexity ratio, with pruned weights on Hugging Face. @arnavmehta007's leap to shrinking trillion-parameter models is his own speculation, not a result. The hardware angle comes from @alexocheema (relayed by @twid): since inference is memory-bound, $/GB/s is the metric that matters, and M5 Ultra at $4.5/GB/s undercuts DGX Spark's $25.5 and DGX Station's $21.1. His own caveat: Apple Silicon's inference software is fragmented across roughly 30 unstable engines with no Dynamo, LMCache, or llm-d, which puts off enterprise buyers.
Patch Plex today, then put an agent on log duty
@c7five's warning is the day's most actionable item: the public Plex chain gives unauthenticated code execution, and the persistence trick is the nasty part. The payload sits in Plex's Codecs folder and reruns on every transcode, so killing the miner does not fix it. If your server was exposed and behind on updates, check that folder for .so or .dll files you did not put there.
@jessethanley, citing the @opencode drama, argues everyone needs "at MINIMUM 1x always on agent scanning logs for weird activity starting today." He runs a @bot security agent over @BetterStackHQ logs several times daily and recommends asking it to scan application logs every 24 hours and push protective patches.
The hype ledger: equity pledges, singularity calls, AI;DR
Separate verifiable signals from posters' framing. On the signal side, @TMTLongShort flags that @JasonDClinton, whom he identifies as Anthropic's CISO, has only 3k followers; @JasonDClinton says he and Brandon pledged 50% of their equity, and that this program was why he left Google for Anthropic before it had a released model. @wassname surfaced the system prompt of Claude's new "you-should-know" plugin for anyone curious how it works.
On the framing side, @kimmonismus reports Anthropic plans 5% of shares in a Public Benefit Fund (subject to shareholder approval, plus 2 points per 50% value increase at the board's discretion, explicitly not a UBI commitment), that Jack Clark said the singularity has been reached, and that "universal high income" is beginning. Treat that synthesis as the poster's own. Same caution for @IterIntellectus relaying Elon Musk's pledge to give one SpaceX share to every kid in America next year.
Culture fallout: @TrungTPhan (via @steipete's "AI;DR") says Amazon's 6-page memo culture is being "nuked" by AI-generated text as employees tune out suspected AI memos. @venaInSV pitches Dan Koe's Human 3.0 framework as an AI life-coach prompt across mind, body, spirit, and career. @0xCodez claims Claude Opus 5.5 reverse-engineered his CarPlay in five minutes and posts a step-by-step guide (his BMW iDrive in the video). @RayFernando1337 recommends following AI educator @HarperSCarroll, citing her Stanford AI degrees and four years building AI at Meta.
Vendor audits and weekend hacks
@GergelyOrosz says Autodesk is copying Adobe's playbook of bullying customers, "penny wise, pound foolish." The quoted evidence is @PermitZIPhvac: a decade-long customer paying $380 per user per month received an audit email despite existing authentication and license assignment, and will transition off Autodesk entirely, taking a nationwide expansion's platform decision elsewhere.
Lighter fare: @aneesbhatti's community got macOS 13 running on a PS5 via Reims vGPU, which @SrLinguicaVIVE met with gleeful disbelief in Portuguese. @dhh shared a Linus Torvalds quote insisting Linux "is NOT some kind of 'social warrior' project." @DevMuzzammil released dumbtv, a no-spying, no-ads, customizable TV setup you configure by pointing an agent at the repo, building on his earlier instructions to enable developer mode over ADB and let Claude or DeepSeek handle the debloating.
Practical Takeaway
The strongest theme is verification. If agents are about to write most of your code, as @ClaudeDevs reports internally and @ibuildthecloud predicts for everyone, the investment that pays off is not fancier prompting but a fast, deterministic feedback harness: e2e and property tests plus automated review loops that catch agent errors before humans see them. If your agents touch the browser, audit cost per task too, since delta snapshots and cheaper models both claim order-of-magnitude savings this week.
Sources
Claude writes about 90% of Claude Code's code. From June to September, merged PRs more than doubled while the share reverted fell by more than 40%. We wrote about how we changed code review to get there, and how to adopt it: https://t.co/9sZe64g3mb
DeepSeek V4.1 Flash 1017 just matched Opus 5.5 on long horizon browser use. 19x cheaper 👀 > 75/100, same as Opus 5.5, almost Astra level > $0.22/task vs $2.35/task > Sol it still SOTA, but it's close The new pareto frontier 🔥 https://t.co/AjefEYpIRd
Did you know the most sophisticated AI systems were actually so chaotic? Yep - this AI meme is actually pretty accurate. “Deep learning” (as mentioned in the X post) refers to neural networks, which are the models that underlie the “generative” AI systems you know about: the text-, image-, video-, and song-generation models. Are you surprised to learn that these systems aren’t straightforward? Does this usage of math make sense to you? Hi, I’m Harper, AI educator, engineer and advisor. ••• My goal is to democratize AI, so everyone can understand it, use it, and make informed decisions about it. Happy you’re here. ••• I have two degrees in Computer Science specializing in Artificial Intelligence from @Stanford, experience teaching PhD-level AI courses there, and 4+ years building AI at @Facebook / @Meta. Follow @harpercarrollai if you want to learn about AI, and let me know in the comments if there’s anything specific you want to learn about.
Tons of bangers and useful advice for undergrads and grad students trying to make sense of how to navigate academia and entering industry in the age of agents in @Tim_Dettmers recent blog. Highly recommend people read it. https://t.co/3VbFSlxqpF https://t.co/b7JOmk2Ibp
How to Reverse Engineer Anything with Opus 5.5 ( Full-Course)
This is a wild approach, @autodesk. We’re a small business paying $380 per user, per month. We’ve been paying customers for over a decade, and we make a deliberate effort to honor the license agreements for every product we use. Your software already requires authentication and an assigned license. What, specifically, prompted this email? Earlier today, I had two conversations on two separate podcast episodes about your direction with Fusion and why cloud-based API access makes sense for agentic MEP workflows. I was making the case for Autodesk. I think the product direction is right based on recent posts from @andrew_anagnost. Our company has been deciding where to invest our development effort: integrate our custom cloud agent harness into Autodesk, build on an open-source platform, or adopt one of the emerging cloud alternatives. That decision will lock in years of internal development and software spending. This decision is also central to our plans to expand @PermitZIP nationwide. An email like this becomes part of that decision. So does the time we have to spend responding to it.For more than a decade, we’ve paid for anywhere from 5 to 30 licenses each year. Autodesk is our single largest SaaS expense. This is a heck of a way to make our next platform decision easier. Starting next week, we’ll begin transitioning away from Autodesk products. We have nothing to hide and will cooperate with the audit. We’ll also be directing our future spending toward companies that earn our confidence as partners. I hope your audit goes well.
yea the great debloating and blocking ads on your tv revolution has begun this is the way: - install adb - get ip address of tv - turn developer mode on that's it claude or deepseek does the rest 🔥 https://t.co/0D1hJciiqe
Here's the multi-agent system driving our software factory right now. Core pieces: - We say what to build (sometimes Linear issues, mostly Slack threads) - Triage agent researches and decides whether to implement or to ask questions (think /grill-me but smarter) - Implementation agent builds, maybe collaborating with us on a spec with a subagent first - Verification subagent tests the implementation agent's work e2e - Code review agent cycles with the implementation agent for a few rounds before human review - Human reviews and ships - Monitoring automation responds to alerts and files issues (our first step to closing the factory loop) There's room to simplify this. We're benchmarking whether we need separate agents for steps like triage vs. implementation, or if a single high-intelligence model can do most of it. The core loop remains the same though. Curious how this compares to factories others are deploying
The usefulness of Amazon’s 6-page memo has been nuked by AI-generated text (employees sit in meetings suspecting text is AI and tune out). Suspicious memos tagged as AI;DR (AI, didn’t read). https://t.co/H1hCGV8elh
Human 3.0 – A Map To Reach The Top 1%
From the community: macOS 13 running on a PS5 with Reims vGPU https://t.co/ta0EMY87mg
@dillon_mulroy Thoughts: https://t.co/gKxqmSn8a7
lots of people like claude's new you-should-know plugin. How does it work? It's main prompt is here https://t.co/pqdsP6envX
Brandon and I pledged 50% of our equity. This was the principal reason that I left Google for Anthropic. The existence of this kind of program was a costly signal that Anthropic wasn’t just talk, at a time when Anthropic didn’t even have a released model. Fund a bright future!
hey. here's a little demo of the most meaningful improvement to my productivity since speech-to-text. https://t.co/ZxOoPIQIrL
The memory unit economics of Apple Silicon are insane. When you realise LLM inference is massively memory-bound (even at high concurrency!), you find that $ per memory bandwidth ($/GB/s) is the most important metric for inference. M5 Ultra: $5499, 96GB @ 1.2TB/s, $4.5/GB/s DGX Spark: $6950, 128GB @ 273GB/s, $25.5/GB/s DGX Station: ~$150K, 252GB @ 7.1TB/s (HBM3e), $21.1/GB/s M5 Ultra has 5.6x better $/GB/s than DGX Spark and 4.7x better $/GB/s than DGX Station. I am surprised Apple didn't increase their prices more (maybe they will so might be a good time to buy...) I expect 512GB M5 Ultra Mac Studio to start at $15899 (following the usual +$25/GB incremental memory price). That's a crazy good price for that much high speed memory. With RDMA over Thunderbolt 5 you get single digit microsecond latency and you can connect up to 5 Thunderbolt 5 cables between Macs for 400Gb/s network bandwidth (same as ConnectX-7). Thunderbolt 5 cables are ~$25 each. The main problem is the software on Apple Silicon is not great. There are about 30 competing inference engines, each supporting a small set of models, and quite unstable. The ecosystem around MLX is immature and limited (e.g. no Dynamo, LMCache, llm-d, ...). In our experience, more serious customers like Governments and Enterprises are put off by that and want something stable with an ecosystem of tools and software to run inference in production.
In addition to this, ask your @bot to scan application logs every 24 hours and actively push patches to protect your app. You’ll be shocked what you find!
We REMOVED upto 10 WHOLE layers from Qwen3-8B-Base using LaCo (Layer Collapse). Upon deleting only 6 layers, the model got 16.7% smaller, keeping 78-94% of its reasoning scores at a 1.68x perplexity ratio. Code and pruned weights are up on our Hugging Face. Here are insights from our research :
Universal high incoming is coming. Anthropic plans to put 5% of its shares into a Public Benefit Fund, subject to shareholder approval, with another 2 percentage points potentially added for every 50% increase in company value at the board’s discretion. The fund could support nonprofits or make direct payments to specific groups affected by AI, but it is not a commitment to universal basic income. Elon Musk supports the idea and is even calling for more direct payments to the general population. Anthropic seems certain: co-founder Jack Clark stated yesterday that we have now reached the singularity. Welcome to the future, friends!