🎧 Listen ADS-FREE: https://podcasts.apple.com/us/channel/djamgamind/id6760446113
Visit our Research Hub at https://djamgamind.com/pdfs
Subscribe to our Newsletter at https://enoumen.substack.com
Summary: In today’s briefing, we investigate “The Recursive Threshold and the Inversion of Web Traffic.” We analyze Anthropic’s alarming white paper revealing that 80% of its internal code is written by Claude, alongside their proposal for a global nuclear-style treaty to pause frontier scale-up. We break down the historic milestone where bot traffic officially eclipsed human internet traffic. We deconstruct the death of “tokenmaxxing” as Microsoft and Snowflake pivot toward highly efficient local architectures. Finally, we cover OpenAI collapsing Codex into ChatGPT alongside its new “dreaming” profile engine, hidden facial-recognition biometric code inside Meta’s smart glasses, and the U.S. government exploring voluntary equity stakes in frontier labs.
Important Topics:
Anthropic Warns of Self-Improving AI: Anthropic publishes data showing that over 80% of its internal code merges are Claude-authored, warning that recursive self-improvement could arrive before institutions are structurally prepared.
Bots Outnumber Humans Online: Cloudflare confirms that automated traffic has inverted the web, with AI agents and web crawlers accounting for 57.5% of total internet activity, leaving human traffic at 42.5%.
The Death of Tokenmaxxing: Enterprise buyers scale back flagrant token expenditures, forcing platforms like Snowflake and Microsoft to launch hyper-efficient, smaller reasoning models and local hardware tools like the Surface Ultra.
ChatGPT Overhauls Memory with “Dreaming”: OpenAI introduces a continuous background optimization engine that transforms disjointed memory facts into a structured, evolving profile of user data.
AI Leaders Group Against Bioweapons: CEOs from OpenAI, Anthropic, DeepMind, and Microsoft sign an open letter pushing Congress to mandate strict customer verification for synthetic DNA and RNA providers.
U.S. Explores AI Ownership Stakes: The federal government enters preliminary discussions regarding voluntary equity distribution from top AI labs, potentially routing proceeds into a household public dividend.
Alibaba Launches Closed Qwen3.7-Max: Alibaba debuts its smartest closed-weight reasoning model, achieving a highly controlled 23% hallucination rate by intentionally declining to answer more than half of its prompts.
Meta Hides Biometric Code in Glasses App: Code analysis reveals Meta has embedded un-activated facial recognition, face-cropping, and biometric encoding models into its smart glasses companion software.
Allen Control Closes $200M Series B: The defense technology firm secures massive backing at a $2.2 billion valuation for its “Bullfrog” autonomous AI-powered anti-drone weapons turret.
🔗 RESOURCES
Stop scrolling through generic “Top 10” lists and head over to the AI Executive Toolkit at https://djamgamind.com/toolkit
Find REMOTE AI Jobs (Mercor): Apply Here
AI & ML TUTOR PRO: https://apps.apple.com/ca/app/ai-ml-tutor-pro/id1610947211
⚗️ PRODUCTION NOTE: We Practice What We Preach.
AI Unraveled is produced using a hybrid “Human-in-the-Loop” workflow.
Anthropic warns that AI will soon be able to improve itself LINK
Anthropic is warning that AI systems will soon be able to build their own successors, a turning point it says “could come sooner than most institutions are prepared for” based on the current pace of progress.
To stop humans from losing control, the company is proposing a global slowdown or temporary pause on AI development, giving alignment research and societal structures time to catch up with the technology.
Anthropic admits a pause would need multiple well-resourced labs across several countries to stop under the same conditions and verify each other, and it points to nuclear-weapons treaties, which took decades, as a model.
Anthropic charts path to self-improving AI
Image source: Anthropic
The Rundown: Anthropic just published “When AI builds itself,” a report on recursive self-improving (RSI) systems — citing internal data on Claude’s coding takeover and cautioning that fully self-improving AI could arrive before institutions are ready.
The details:
Anthropic noted that RSI is not here yet or even inevitable, but said that Claude is advancing AI development “faster than we thought”.
More than 80% of Anthropic’s merged code was Claude-authored as of May, with engineers pushing 8x as much code per day in Q2 2026 as in 2024.
“Each new version of Claude could be built by the version before it, without human involvement,” co-author Jack Clark wrote of where the trend leads.
OpenAI flagged the same loop this week in its “Democratic Governance of Frontier AI” blueprint, pointing to RSI’s first sparks in today’s systems.
Anthropic said it would slow or pause frontier AI if peer labs did too, and plans policy talks in the coming months to discuss research, systems, and scenarios.
Why it matters: Anthropic and OpenAI aren’t alone in feeling the RSI, with labs like MiniMax saying its M2.7 model helped build itself and new startups dedicated to the self-improvement loop popping up everywhere. The unknowns of RSI are scary, but it’s also hard to fathom a feasible pause scenario that hinges on global coordination.
OpenAI’s memory overhaul lets ChatGPT ‘dream’
Image source: OpenAI
The Rundown: OpenAI just introduced a new memory update within ChatGPT centered around “dreaming,” a background system that turns past chats into a running, category-sorted profile of who you are for better personalization and evolving context.
The details:
ChatGPT now keeps a running written summary of each user, grouped into areas like travel, hobbies, and work, replacing the previous list of one-off facts.
Users can review memories, make corrections, add details, or ask ChatGPT not to bring up certain topics, with memory automatically updating over time.
OpenAI says factual recall rose from 41.5% to 82.8% in its evals with dreaming, while preference-following climbed from 31.4% to 71.3%.
Dreaming is rolling out to Plus and Pro users in the U.S., with Free and Go, and more countries getting the upgrade over the next several weeks.
Why it matters: Memory is one of the stickiest AI features in theory, but it’s been long overdue for an upgrade. Sam Altman frequently talks about the future of hyper-personalized AI, and dreaming may become a big part of forming that continuity and proactivity for users — and keeping them from switching to rivals in the process.
Rival AI labs unite behind bioweapons risks
Image source: Open letter to Congress
The Rundown: CEOs of OpenAI, Anthropic, Google DeepMind, and Microsoft just signed an open letter pressing Congress to make synthetic-DNA sellers vet every buyer/order, warning that AI can now enable bad actors to create bioweapon designs.
The details:
Signees include Sam Altman, Dario Amodei, Mustafa Suleyman, Alexandr Wang, and Demis Hassabis, alongside DNA-synthesis industry leaders.
The letter said, “AI systems now outperform PhD-level virologists… about highly technical lab procedures in their own domains of expertise.”
The signers urge Congress to make U.S. synthetic-DNA and RNA sellers screen orders, verify buyers, and log sales to keep dangerous sequences traceable.
They also warned that “knowledge barriers which have historically prevented bad actors from obtaining biological weapons will meaningfully erode.”
Why it matters: Like RSI, biological weapons concerns have been outlined for years as one of the steps up the ladder of AI improvement — and this one’s concerning enough to even unite Altman and Amodei. It’s clear that laws and regulations are going to have to adapt quickly to a changing AI-accelerated world, but the question is if they can.
What ChatGPT will do with its Codex superpowers
OpenAI is collapsing the wall between ChatGPT and Codex, folding them into a single platform so users no longer have to toggle between apps, and opening Codex to far more people than the coders who first relied on it.
Instead of making you choose between using the fast-growing Codex agent and ChatGPT, which has become the company’s biggest brand used by nearly a billion users, the company decided to take all of Codex and put it inside ChatGPT, Alex Embiricos, head of enterprise product, explained in a Codex workshop for press in New York City.
The result: easier access to Codex’s powerful capabilities.
“It’s been really fun working with software developers, because they’re an audience that wants to try a lot of new things,” said Embiricos. “We’re now at the second phase, where we have these incredibly useful agents that are actually useful to do anything you can do on your computer. And now our goal is to bring this capability to everyone.”
This update will also improve interoperability, Embiricos explained, as users on a computer or phone will have access to the same Codex capabilities directly within ChatGPT.
Also this week, OpenAI launched six new role-specific plugins that target roles beyond traditional software development, such as creative production, sales, and public equity investing, highlighting the push to bring Codex to a broader audience.
“For us, this is really core to our mission. We really care about making sure that the best technology is broadly available to everyone,” said Embiricos.
If, like me, you’ve been skeptical about what Codex can actually do for non-coders, the workshop I attended shed some light on a few ways everyday users can put it to work. Here’s what I found genuinely useful:
Sending messages: By connecting your email, Slack, or another messaging platform, you can describe what you want to say, and Codex will draft it, locate the recipient’s information, and send it with a single approval from you.
Calendar briefings: Connect your calendar via a plugin, and you can ask Codex questions about upcoming events, or have it surface specific information on demand.
Daily automations: Once you’ve combined the right plugins for a given workflow, you can instruct Codex to run it automatically at whatever cadence works best for you. For example: Every morning at 8 AM, give me a summary of today’s meetings and the most important action items from my inbox.
The tokenmaxxing era is over before it started
AI customers may be starting to pinch their pennies, and tech giants are taking notice.
At both Microsoft Build and Snowflake Summit this week, efficiency stood out as a prevailing theme in the announcements of these enterprise tech giants. It may signal that the compute costs that are crunching AI builders are starting to add up, and flagrant spending fueled by sky-high expectations may be starting to come back to earth.
“I think if you read about OpenClaw’s founder, Peter Steinberger, and how many millions of dollars worth of tokens that he’s using, it doesn’t necessarily correlate to an output,” Rob Ferguson, VP of technology and strategy at Fireworks AI, told The Deep View this week. “People are starting to really think about what the outputs of their AI are.”
In short, the era of “tokenmaxxing“ may be over. Or, at least, the definition is changing, said Ferguson. Rather than focusing on eating up as many tokens as their competitors, enterprises are starting to think about how to squeeze as much as they can out of the tokens they use.
Several of the product releases in San Francisco this week back up that shift:
Snowflake’s new Cortex Training system, which allows enterprises to customize open-weight foundation models, is marketed specifically as being faster and less expensive. Additionally, Snowflake’s new Adaptive Compute addresses cost efficiency at the infrastructure level by automatically calculating the best use of compute and software resources in real time.
Microsoft’s new models also reflect a desire for efficiency, with its first reasoning model sitting at 35 billion parameters (compared to the latest trillion-parameter models that OpenAI and Anthropic offer) and built specifically for efficiency and low-token cost.
The company is even targeting efficiency on the hardware side, debuting both the Surface Laptop Ultra and the Surface RTX Spark Dev Box, which can run powerful models locally and drastically reduce token costs. Jatinder Mann, partner director of product management at Microsoft, told The Deep View that these devices aim to provide “unmetered intelligence,” reducing cloud costs by enabling local models to handle routine tasks. “There are a lot of routine things that don’t necessarily need a cloud model,” Mann said.
The next step enterprises should take is questioning whether a task requires AI at all, Raj Ramanujam, VP of Global Alliances and Cloud at Dynatrace, told The Deep View. Every agentic task, every prompt, every tool call racks up the bill. It’s why every potential AI implementation should start with a “problem statement,” he said, identifying exactly what challenge they’re trying to solve or task they’d like to automate.
“There are some things that you can automate without touching AI in the normal course of how you program it,” said Ramanujam.
5 things to expect at WWDC 2026 LINK
WWDC 2026 is shaping up to be a pivotal event for Apple, with the keynote expected to focus heavily on Siri, a wave of Apple Intelligence updates, iOS 27 changes for the iPhone Fold, and possible new hardware previews.
Siri is rumored to gain chatbot-like abilities and a dedicated app living inside the Dynamic Island, while reportedly running on Google’s Gemini after Apple struck a deal to power its assistant with the rival AI.
iOS 27 is expected to be a Snow Leopard-style performance release, with a new Visual Intelligence feature across core apps, boosted image generation in Photos, AI-built Shortcuts, and possibly opening streaming protocols to apps like Google Cast.
Qwen3.7-Max Adds Speed and Power
Alibaba updated its flagship large language model for long-running agentic work, pushing it into the top rank among LLMs built in China.
What’s new: Alibaba positions Qwen3.7-Max as its preferred model for text-only work like coding and scientific discovery. Like other top-tier Qwen models since late 2025, its weights are not open. (Simultaneously Alibaba released the multimodal Qwen3.7-Plus-Preview.)
Input/output: Text in (up to 1 million tokens), text out (up to 64,000 tokens, 208.3 tokens per second)
Features: Reasoning, tool use, prompt caching, native compatibility with OpenAI’s and Anthropic’s API specifications, ability to retain reasoning text across turns
Performance: Ranks seventh on Artificial Analysis Intelligence Index
Availability: Free via Qwen Chat (account required); API via Alibaba Cloud Model Studio $2.50/$0.25/$7.50 per million input/cached/output tokens
Undisclosed: Parameter count, architecture, training data and methods
How it works: Alibaba described Qwen3.7-Max’s reinforcement-learning approach at a high level. The approach separates three components that Alibaba says are typically coupled in agent training: the task to be performed, an agentic harness that calls tools, and a verifier that decides whether the system succeeded. Alibaba trained the model on many combinations of task, harness, and verifier to prevent it from learning tricks specific to a single setup.
Performance: Qwen3.7-Max trails the top tier of reasoning models on the Artificial Analysis Intelligence Index, just behind leading U.S. models from OpenAI, Anthropic, and Google. It excels at delivering correct output partly by declining to respond more often than peers.
On the Artificial Analysis Intelligence Index, a composite of 10 tests of economically useful tasks, Qwen3.7-Max set to reasoning (56.6) ranks fifth or seventh depending on the reasoning levels of various models. It’s behind Gemini 3.1 Pro Preview set to an unspecified level of reasoning (57.2) and ahead of Google Gemini 3.5 Flash set to high reasoning (55.3). Running those evaluations consumed roughly 97 million output tokens, well above the average 35 million tokens.
On AA-Omniscience, an Artificial Analysis measure of factual knowledge that rewards correct output, penalizes incorrect output, and doesn’t count abstentions — which helps to distinguish between models that do and don’t acknowledge the limits of their knowledge — Qwen3.7-Max set to an unspecified level of reasoning ranked sixth (14), well behind Gemini 3.1 Pro Preview set to reasoning (33) and ahead of Claude Sonnet 4.6 set to max reasoning (12). Qwen3.7-Max’s 23 percent hallucination rate was the lowest among frontier models tested, but did so partly by declining to respond to more than half of the prompts.
In Artificial Analysis’ measure of output speed, Qwen3.7-Max set to an unspecified level of reasoning (208 tokens per second) tied for third place with Gemini 3.5 Flash — reasoning unspecified — just behind GPT-OSS 120B (313 tokens per second) and GPT-OSS 20B (238 tokens per second).
Yes, but: Although Alibaba touts Qwen3.7-Max’s agentic capabilities, the claim is based on an internal test that is not yet validated by independent benchmarks. The model autonomously optimized an attention kernel on hardware it had not encountered during training. In 35 hours, it made 1,158 tool calls and ran 432 kernel evaluations (test runs of candidate code). The resulting code ran roughly 10 times faster than a standard reference implementation. Artificial Analysis has not yet tested Qwen3.7-Max on its benchmark of long-running agentic tasks.
Behind the news: Qwen3.7-Max continues Alibaba’s shift from open to closed models. In addition to Qwen3.7-Max, Qwen3.6-Max-Preview and Qwen3.6-Plus have closed weights, while the weights for the less capable Qwen3.6-27B and Qwen3.6-35B-A3B are freely available. At the same time, Alibaba started charging for access to Qwen Code, a command-line coding tool. These changes follow turnover in the Qwen team’s leadership and suggest that Alibaba aims to leverage its top-tier models to produce revenue rather than maximize its reach.
Why it matters: Qwen3.7-Max is the smartest Chinese LLM, judging by the Artificial Analysis Intelligence Index, and it’s the third-fastest overall.
US weighs government stakes in AI firms LINK
The US government is in early talks with major AI companies about taking ownership stakes in the firms building frontier AI, with shares voluntarily handed over rather than purchased, according to a NOTUS report.
Sam Altman first pitched the stake idea to Donald Trump in early 2025 and has raised it again in recent weeks, with one option sending proceeds to a dividend paid to every American household.
Critics quoted in the report warn the setup makes the government both shareholder and referee of the same companies, while polling cited shows 55% of Americans think AI will do more harm than good.
Meta hides face-recognition code in smart glasses app LINK
Meta has quietly placed face-recognition code for its smart glasses inside a phone app already downloaded by millions, according to a WIRED analysis that outside experts independently reproduced on their own devices.
Three AI models tied to a feature called NameTag now sit on customers’ phones, with one detecting faces, one cropping them, and a third encoding them into biometric data.
Facial recognition is not switched on, but an Electronic Frontier Foundation researcher who tested the code said it appears partly functional, while Meta spokesperson Andy Stone claimed on X the “feature doesn’t exist.”
Cloudflare CEO says bot internet traffic has overtaken humans LINK
Cloudflare CEO Matthew Prince says bot traffic has now overtaken human traffic on the internet for the first time, arriving earlier than his prediction of late 2027 due to fast-growing agentic traffic.
Cloudflare’s data shows that between 52 and 62 percent of daily internet traffic comes from bots, with the past week averaging 57.4 percent bots and 42.5 percent humans, including search crawlers and AI bots.
By country, Gibraltar leads with 92.1 percent bot traffic, followed by Singapore at 76.3 percent, Iran at 76.2 percent, Ireland at 72.8 percent, and the Netherlands at 68.8 percent, driven by AI agents scraping and acting for users.
AI now designs antigens:
A team from the University of Cambridge developed what they’re calling a “universal vaccine” after developing a “super-antigen” that could provide an “all-in-one” solution to future outbreaks. The team fed an AI model every known genetic sequence from a sub-genus of coronavirus known as “Sarbeco.” The model then developed an antigen with common features, that could — at least in theory — help block or even prevent infections from the entire viral group. So far, the results have only been tested on a relatively small group of volunteers at two UK medical facilities, but it appears to have triggered a positive immune response against both SARS and COVID. If AI helps to prevent a whole swath of dangerous illnesses, will everyday Americans start trusting it more? It’s at least a start!
Bots are more online than people:
According to new data from Cloudflare, excitedly tweeted by CEO and co-founder Matthew Prince, “bots have now passed human traffic online for the first time in the Internet’s history.” That obviously includes old fashioned internet bots, like website crawlers and search indexers, but Cloudflare suggests this massive recent surge is all thanks to AI agents. That includes agents performing multi-step tasks on behalf of human users, but also bots scraping and indexing content for future AI training. Which, let’s be fair, is probably a decent-sized chunk of that total. As of now, CF estimates bot traffic is behind a whopping 57.5% of total internet activity, compared to 42.5% for us flesh bags.
Allen Control raises $200M Series B:
The robotics defense tech company (HQ’d right here in Austin, Texas!) develops autonomous anti-drone weapons systems. (Or “robotic turrets,” as some headlines have put it.) Their flagship product to date is Bullfrog, an AI-powered weapons station that helps legacy firearms more precisely engage with targets, while also providing defense capabilities (like passive sensing that helps troops maintain covert positions). They’ve raised a $200M Series B at a $2.2 billion post money valuation, led by Disney vet Kevin Mayer’s Smash Capital.
