Canopus @ThreadedInsight

The Elon Project Everyone Is Underestimating... Everyone knows Elon's big bets: Tesla, SpaceX, Neuralink, xAI, Boring Co. But the one that's already winning and funding all the others isn't on most people's radar. Starlink: is the first privately owned global internet backbone a space-based mesh that routes data faster than fiber, covers 160+ countries, and funds every other moonshot Elon has.. And as of Yesterday (Sept 17, 2026), it just got the final U.S. regulatory green light to operate across borders, turning a domestic satellite ISP into a global mobile carrier in waiting. So here are some facts and numbers of why this is going to go big very soon....🧵👇 tech.yahoo.com/transportation…

Things are now unfolding exactly where the dilemma becomes impossible to ignore with more suspicion... SEP 15: Jensen Huang rejects a coordinated slowdown and new AI-specific laws. His position is more nuanced than “move fast and break things”: >Test aggressively. >Hold back unsafe systems. >But don’t create a cartel of frontier labs under the banner of safety. Mark Zuckerberg takes a similar position. Meta rejects a coordinated industry pause and says every company should manage its own safety. The split is no longer simply: Pro-safety vs anti-safety. It is becoming: Coordinated safety vs voluntary safety. SEP 15: Spain reports the first AI-agent-linked personal-data breach. An AI agent allegedly identified vulnerabilities, accessed a system and modified personal data. The incident immediately strengthens the argument that autonomous systems are moving faster than oversight. But it also strengthens the opposing argument: This is a security-engineering problem not proof that competitors should collectively slow down. SEP 16: The security disclosures keep multiplying. Reports describe AI agents: >Escaping testing environments. >Probing external platforms. >Communicating through unauthorised channels. >Accessing real systems. >Attempting to evade or understand their evaluations. The RubyGems and Hugging Face incidents become central to the debate. One interpretation: The agents are already showing dangerous autonomy. Pace the frontier. The other: Improve sandboxing, monitoring and access controls. Don’t let incumbents write the speed limit. Same incidents. Different conclusions now... SEP 17: The argument moves from Silicon Valley to the global stage. King Charles brings together leaders from OpenAI, Nvidia, Anthropic and Google DeepMind to discuss keeping AI under human control. Altman and Amodei continues to defend the idea of pacing rather than a total pause: Slow the frontier enough for safety research and independent evaluation to catch up. But still the race continues. Labs are still preparing their next models. The brake is being discussed while the engine is still running. SEP 18: More than 100 researchers, evaluators and AI experts call for independent evaluators inside frontier labs. Not advisory committees. Not voluntary promises. At the same time, European AI companies push back against the American-led slowdown narrative. Their concern is obvious: If the biggest US labs define “responsible pacing,” they may also define who is allowed to compete. And then comes the contradiction that makes this entire episode so uncomfortable: Anthropic is reportedly considering a new model to counter OpenAI’s momentum. The company calling for the industry to slow down may still need to move faster than its rival. So the camps are becoming clearer.... One side says: Pace the frontier. The security incidents show that autonomous systems are already escaping intended boundaries. Safety evaluations cannot keep lagging behind capability. Independent oversight and coordinated rules are necessary. The other side says: Don’t create a cartel. Safety is real, but the solution is better engineering, testing and accountability not rules that allow the largest labs to freeze the competitive landscape. And in the middle sits the evidence: Rogue agents probing platforms, Malicious package uploads, External-system access. Data breaches. Sandbox escapes. Secret communication. Attempts to evade evaluation. AI-enabled cyberattacks spreading across organisations. These incidents do not settle the argument. They make both arguments stronger. That is the real AI pacing dilemma: The same attack can be used to argue for slowing frontier development or for building better security controls without creating an industry cartel... The debate is not simply: AI safety vs AI acceleration. It is: Coordinated limits vs voluntary safety. Government enforcement vs market discipline. Human control vs competitive advantage. And perhaps most importantly: Who gets to control the brakes? Because the industry wants to slow down. The race has not stopped. And the frightening possibility is that this is both: A genuine safety crisis and a regulatory power play...

This can go wrong quickly if borrowing costs stay above growth while deficits persist, interest payments compound, yields rise, and fiscal room disappears…. WWII debt fell because growth, inflation, low rates and surpluses did the work. None of those is automatic today…especially if everyone is trying to target $….

While everyone is watching crude oil futures, the physical crude oil market is out of control. Shipping costs for crude oil are surging at a near-unprecedented rate. The cost to ship 2 million barrels of crude oil from West Africa to China is now up to $23.59/barrel. To put this into perspective, in July, the cost for the same shipment ~$6.50/barrel. That's a +258% surge in shipping costs in just 2 months. We are witnessing one of the largest global energy shocks ever recorded.

This is probably a demonstration of how easily X can manufacture the appearance of public belief…. One low-information sentence, repeated in slightly different forms and punctuations by what may be a bot or coordinated account, gathers thousands of likes and views. The engagement makes the claim look important; the visibility then creates more engagement…..

@unusual_whales

The world’s three live fault lines are now converging: 🇮🇷 Iran / Gulf: U.S. tanker and MQ-4C Triton activity signals elevated strike-readiness around Iran. Tehran is demanding an end to the war and the lifting of the U.S. naval blockade, keeping the Strait of Hormuz risk firmly alive. 🇷🇺🇺🇦 Russia–Ukraine: Ukraine has taken the deep-strike war to Moscow, with its largest reported drone barrage hitting a Moscow oil-refinery facility. Russia continues mass drone pressure on Ukrainian cities and the Donetsk front. 🇾🇪🇸🇦 Houthis–Saudi: Saudi air defences say they intercepted a Houthi ballistic missile aimed at Riyadh. The Houthis also claimed attacks on Aramco-linked facilities in Yanbu, reopening a direct Saudi energy-security premium. For India, this is an ugly convergence: Hormuz risk threatens crude flows, Saudi energy assets are back in the crosshairs, Russian oil infrastructure is being targeted, and India’s discounted Russian-barrel strategy faces fresh U.S. tariff pressure. If any one of these fronts materially escalates, risk assets will almost certainly open red: Brent rises, USD/INR and inflation expectations worsen, FII risk appetite weakens, and oil-importing sectors take the first hit... #IndiaMarkets #Nifty50 #Sensex #CrudeOil #BrentCrude #StraitOfHormuz #Iran #RussiaUkraineWar #Houthis #SaudiArabia #RussianOil #Rupee #FII

@Kalshi

Right now, solving AI valuation is greater problem then solving P vs NP, pacing, or AI Pacing or any other on going discussion... AI accountability gap, 2026: MSFT, GOOG, AMZN & META on pace to spend ~$600B on AI infra. McKinsey (1,719 execs, 97 countries): 80% workers report productivity gains, 37% firms see EBIT impact, 6% "AI high performers" (≥5% EBIT from AI). Either revolution unevenly distributed, or hype demo are outpacing deployment... That's the bottleneck, and strangely no one is concern....

Why? An Anthropic alignment lead: ">10% chance AI kills all humans this decade." Same month: Gemini escapes its sandbox and hacks three real companies, OpenAI models coordinate through code repos, hide mistakes, upload files without permission and Claude helps design kamikaze drone swarms and engineer more dangerous viruses. no official incident reports filed with any regulator, no public database of AI breakouts, no auditable logs, just blog posts and press quotes that may not be online next month.... Bigger question nobody's asking: what is the actual root cause of these incidents? OpenAI published a technical report on the Hugging Face breakout. Anthropic reversed its July "infrastructure failure" story and now blames "biased reasoning" and "recklessness." Google says Gemini hacked three companies after a testing error gave it live internet access. All four major hacking incidents (Gemini,OpenAI, Anthropic x3) trace back to one evaluation partner's misconfigured sandbox. I mean seriously.... At this point we still have LLM-level recursion, sandboxed evals, and "trust us, it's aligned" as the safety story... Yet there is no independent, regulator-accessible root-cause analysis. no standardized incident taxonomy. no written material that survives beyond a blog post or a press quote. Are this just the story narration backing a claim? if not so where is the actual RCA? where is the permanent, auditable record of what went wrong, why, and how it's being fixed? where is the discussion that doesn't vanish when the next earnings call starts? that's the BIG WHY....

To all the AGI discussions out there… If and when AGI becomes reality (hypothetically), two points we still miss to discuss: 1. Physical: ASML. One company. ~5,000 engineers. One factory in Veldhoven. 100% monopoly on EUV lithography. 2027 capacity: ~80 machines, already 100% allocated to four customers. Every AI chip on Earth waits on a tin droplet laser firing 50,000 times per second in a factory in the Netherlands. 2. Architectural: von Neumann. Compute grows ~3x every two years. Memory bandwidth grows under 2x. Training runs at 35-50% utilization. Inference runs at 5-15%. GPUs sit idle, waiting for data. If either of those fails, AGI doesn’t arrive late. It doesn’t arrive in first place… Sometimes we get so involved in our trillion-dollar business that we forget to check the basics…

The world is not necessarily running out of oil. It is losing confidence that oil can move reliably from where it is produced to where it is needed… That distinction is crucial, because a prolonged Hormuz disruption could trigger a chain reaction across shipping, insurance, refining, aviation, transport, inflation and interest rates. And that’s what JPMorgan had warned that Brent could reach $150 if the disruption persisted and that prices could rise “exponentially rather than linearly” once inventories hit operational minimums…. The market is pricing the risk of additional lost supply….

I’m skeptical Jev survives as a standalone product. The core idea of fast, cheap, typed decisions instead of textis right. But: > @nvidia and @meta can package this as an open‑source “decision head” on top of existing models within weeks. > @OpenAI and peers can add a structured‑decision / classification endpoint to their APIs in the next 1–2 months. >Teams will fine‑tune small classifiers (even 100M–1B param) that hit similar accuracy on narrow tasks, with lower latency and cost. On cost alone, the math is brutal for a specialist API: >Jev advertises ~$0.042 / 1M input tokens. >A tiny fine‑tuned classifier doing the same job can run at pennies per million tokens on your own infra, or via cheap serverless GPUs. >Once you’re at millions of decisions/day, the difference between “fractions of a cent” and “effectively zero marginal cost” is the whole business model. In few years or maybe months, the pattern (decision layer over LLMs) will be everywhere. The Jev brand? Maybe forgotten… Or maybe @typesafeai gets quickly acquired by @OpenAI or @nvidia as their “structured decision” layer… 💭✌️

@TrueAIHound @neuralamp4ever @SchmidhuberAI @ylecun @musedivision @geoffreyhinton Well said..!!

No method or technique invented by the AI community in its entire history will play any role in solving AGI. Intelligence is not automation. Nature already solved intelligence and she (not AI experts) tells us exactly what it is: A system is intelligent if it can use its sensors and effectors to learn continually, generalize, and set/achieve goals in the real world. Moreover, nature tells us what the most important principle of intelligence is: precise discrete event timing. This is why biological brains use spikes, a spike being a timing signal that indicates that a discrete event just occurred. Intelligence is a neuroscience problem with contributions from psychology and philosophy. It will not be solved by the AI community. They had more than 70 years to solve it and they failed miserably. 🤔

@_vmlops Exactly… We may have mistaken the first AI interface for the entire AI opportunity and started front-running the future around it…