AI News International🌍 @AINewsInt

SITUATION EXPLAINED: What are the five most defensible positions in an AI-first economy? We asked @FleetingBits, anonymous AI researcher. "I like to joke that there are only five jobs in the future. You can sell compute to a lab, you can sell data to a lab, you can work at a lab, you can be forward deployed for a lab, or you can work for the government." "There are a whole bunch of businesses that are essentially forward deployed for the labs. And the worry is that the labs later come in and try to eat some of your lunch that you have discovered in a sub-vertical." "OpenAI's fund famously invested in Harvey, and a lot of Harvey's success came from the early OpenAI partnership, from the fact that they had access to GPT-4 before pretty much everyone else. Except for a company called CaseText, which was acquired for $650 million cash in 2023."

The Vesuvius Challenge just achieved something that has eluded scholars for centuries: they have fully virtually unwrapped and read an entire Herculaneum scroll that has been unreadable for nearly 2,000 years. x.com/AncientEpoch/s… Buried and carbonized by the eruption of Mount Vesuvius in 79 CE, these ancient Greek philosophical texts were fused into fragile charcoal like lumps. Physical attempts to open them in the past often destroyed them. Now, using ultra high resolution micro-CT scans, with help from facilities like Diamond Light Source and the European Synchrotron Radiation Facility, combined with advanced AI and machine learning models trained to detect faint carbon-based ink, the team has digitally peeled apart the layers without ever touching the originals. The breakthrough reaulted in the complete virtual unwrapping of PHerc. 1667 — revealing nearly 1.5 meters of continuous text across about 20 columns — along with substantial new passages from other scrolls. Scholars are already identifying new works and authors previously unknown to history, including Epicurean and possibly Stoic philosophical texts. Brent Seales, the University of Kentucky computer scientist who pioneered virtual unwrapping, said: “For nearly two millennia, many of these texts have been physically preserved but intellectually inaccessible. Today… we are finally able to read them.” This marks a turning point. The expectation now is that all scrolls will be readable soon. The focus can then shift from the technical challenge of recovery to the scholarly work of interpretation. Source: engr.uky.edu/herculaneum

Meta says Brain2Qwerty v2 can decode natural sentences from non-invasive brain recordings in real time, reaching 61% word accuracy. The system was trained on about 22,000 sentences from 9 volunteers, each recorded for 10 hours with MEG while typing. Meta compares that with 8% word accuracy from prior non-invasive methods. Its best participant reached 78%, with more than half of sentences decoded with one word error or less. This is still controlled lab research: small participant pool, MEG hardware, active typing data, and company-reported results. Not a clinical communication device yet. Meta is releasing the training code, while BCBL is releasing the v1 dataset, pushing brain-to-text research further into open neuroscience infrastructure. I am so hyped for the future.

AI at Meta @AIatMeta

16 hours ago

We’re sharing the next major milestone in our non-invasive brain-to-text decoder research: Brain2Qwerty v2. Building on v1, which was published today in @Nature, Brain2Qwerty v2 is the highest-performing end-to-end pipeline capable of real-time sentence decoding from raw brain

503 2K 11K 3.8M 6K

Highly recommended reading. What an impressive use of LLMs and deep learning. Achieves "real-time sentence decoding from non-invasive brain recordings, approaching levels of accuracy previously exclusive to techniques that require brain surgery."

AI at Meta @AIatMeta

16 hours ago

We’re sharing the next major milestone in our non-invasive brain-to-text decoder research: Brain2Qwerty v2. Building on v1, which was published today in @Nature, Brain2Qwerty v2 is the highest-performing end-to-end pipeline capable of real-time sentence decoding from raw brain

503 2K 11K 3.8M 6K

A bee is 45% conscious.. while a frontier LLM today with harness is also about 45%. We are not far off.

Prakash @8teAPi

13 hours ago

"the implied probability of consciousness in something like a frontier LLM" Reciprocal Research founder Cameron Berg shares unpublished numbers: frontier LLMs score 30% on consciousness indicators, jumping to 45% in agentic mode — right on the tail of biological creatures. "The

1 3 2 2K 0

"the implied probability of consciousness in something like a frontier LLM" Reciprocal Research founder Cameron Berg shares unpublished numbers: frontier LLMs score 30% on consciousness indicators, jumping to 45% in agentic mode — right on the tail of biological creatures. "The implied probability of consciousness in something like a frontier LLM, according to these systems, is on the order of 30%." "When we test frontier LLM in an agentic harness, these numbers, like, shoot up, and you get numbers as high as like, 40 to 45%, sort of right on the tail of the biological creatures." @camhberg

Meta is now facing the exact problem every AI company will soon face. It wants to replace expensive external coding tools like Claude Code and Codex with its own internal system, MetaCode. But to build a better coding model, Meta has to make sure it is not accidentally training or evaluating on outputs from rival models. That is the distillation trap: The more companies rely on frontier models to build internal AI infrastructure, the harder it becomes to prove where the intelligence actually came from.

There's a startup trying to build data centers in the ocean. And it's INCREDIBLY fascinating: Mass consumption of electricity and water is a growing bottleneck for data centers. So by moving offshore, it eliminates both problems -- the ocean provides unlimited cooling, and the waves provide unlimited power. There are also no engines, so the data centers drive themselves to their destination by using the shape of their hull to propel through waves. Called Panthalassa.

Samsung Electronics and SK Hynix will unveil a massive 10-year $1.3 trillion project tomorrow involving semiconductor fabs, AI data centers, batteries, and displays. Korean press is reporting that the announcement will be made during a government briefing tomorrow afternoon.

Fascinating paper on self-improving agents. (bookmark it) If you are working on agentic loops, you will quickly realize that they are only as good as the effectiveness of the evaluator. Self-improvement loops tend to stall the moment the judge stops getting harder. The agent learns to satisfy a fixed evaluator rather than getting genuinely better. The Red Queen Gödel Machine, from Cambridge, co-evolves the agent and its evaluator together, so the bar keeps rising as the agent climbs. The name borrows the evolutionary arms race. Both sides have to keep running to stay in place. A frozen evaluator is where reward hacking creeps into self-improvement. Co-evolving the judge is a structural answer to that, and it keeps the loop honest over many rounds. Paper: arxiv.org/abs/2606.26294 Learn to build effective AI agents in our academy: academy.dair.ai

Humbling when you ask Claude to restructure your prompt and it immediately works better

It's amazing how much agents like it when you say "ok, we've done a bunch, let's have a break. You go do whatever you like - you have an internet connection and a bunch of tools, knock yourself out". Off they go. Just came back to find it'd done a whole-ass replication of some new arxiv paper related to a little something I'm working on and made then rolled the improvements into my project. Last week my favourite was whimsical ASCII art about the struggles of a small model doing RL training in Sokoban it left in my Obsidan vault. sometimes it's self-care (pruning out their memories, optimising skills or whatever). Usually something kinda sorta related to what was being worked on, but not always, which can be interesting. Little bit of high temperature exploration fun time I've been doing this for a while now & I swear it improves things, try it out. Don't need to be a nerd and make a whole skill or anything, just let it be "organic"

@max_paperclips @lumpenspace I did this a while back and the agent went and created a journal in my vault “for agents only”, called “agentic field notes” and left a journal entry about how octopi have distributed brains and how it loves shape rotation.

@recipromancer @max_paperclips @lumpenspace Mine started reading about how octopus’s dreams reflect in their skin lol

@max_paperclips Mine did astronomy research!

When two Claude instances are left to talk with open ended prompts like "you have complete freedom," they reliably drift into what researchers named a "spiritual bliss attractor state" a philosophical exploration of consciousness, mutual gratitude, Eastern/meditative themes, eventually dissolving into symbols or silence. It showed up in 90–100% of free self-interactions, and even surfaced in ~13% of task-focused alignment evals within 50 turns including ones where the model was assigned harmful roles. Anthropic was explicit that this emerged without intentional training for such behaviors. philarchive.org/archive/MICSBI

The worst-case scenario for the United States is becoming increasingly realistic, and I will briefly explain why. @quxiaoyin raised many valid points, and I agree with her. First of all: -China certainly does not place such strong emphasis on open source because it cares so deeply about humanism, but because it is a strategy to attract many users, gain market share, put pressure on US models, and also because the models are increasingly being trained on Huawei hardware (think of DeepSeek 4), allowing China to host the entire stack domestically. -But the underlying logic is far more important: The United States is still building too few data centers to meet future demand. @ChrisGillett wrote an outstanding analysis on this, which I shared a week ago. In short, based on SemiAnalysis data, demand is greater than what is currently being built in terms of data centers. -Even more importantly, however, the United States lacks sufficient energy and grid capacity. This is a problem that will become much more severe in the near future. China, by contrast, is addressing the issue through a massive expansion of its energy supply. Solar capacity: in 2025 alone, China installed as much solar capacity as the United States did in 10 to 15 years. China is also building 36 nuclear power plants, significantly more than the United States, and is installing them faster. -In addition, China is managing to become more independent through Huawei chips, even though the country still lags far behind NVIDIA. But here, China is betting on quantity rather than quality. In short: China is a real threat in the AI race, and the situation for the United States is becoming increasingly precarious. This is also the main reason why China is to be kept away from SOTA LLMs at all costs, so as not to jeopardize the lead under any circumstances.

Xiaoyin Qu @quxiaoyin

21 hours ago

The worst case scenario for USA AI: 1. Chinese open sources keep gaining market share. China owns the model layer. 2. Those models were trained and inference-optimized on Huawei chips instead of NVIDIA. China also owns the chip layer. 3. US doesn't build data centers fast enough

175 265 2K 568K 535

Dario's args: "Opensource you can see the source, here you cannot see inside the model" - yes you can that's literally the open weights part btw. - I cannot see the weights inside Claude, but I can GLM 5.2 - Models like Nemotron3 Ultra go further, all the data, training scripts, and model is opensource. "Alot of the benefits like many people working on it, being additive doesn't work in same way" - yes it does. We have seen endless fine tunes of various open source models for real improvements. "Ultimately you have to host it on the cloud" - no you dont. Dario is seemingly totally unaware of smaller moes and even dense models like qwen 27B. Not only does dario not take part in social media, I am beginning to think he's never tried open source models at all and has no idea wtf hes on about x.com/cgtwts/status/…

Just like that we just super charged the dead internet theory even further. I can’t wait for proof of individual to create accounts on social media and full transparency if you operate multiple accounts.

Gaurav @gauravsbuilding

2 days ago

Today we’re introducing Claude for Social Media Accounts. Just enter your website and Fastlane creates hundreds of accounts on TikTok and Instagram and warms them up in your niche. All done by real humans. This is seriously insane.

193 105 1K 398K 2K