Canopus

#Tech by Edu #Finance by curiosity. Research on #AI, skeptical of every #Hype Cycle. X-Observer → #Reply & #Posts that matter. #Unfiltered Takes & #Articles.

[Canopus @ThreadedInsight]

Your AI bill is lying to you, Tokens are only 8-27% of what you actually pay. Welcome to "Tokenomics" the emerging playbook for tracking, allocating, and actually controlling AI spend before it controls you. 🧵 https://t.co/aBnshLyaZJ

[Canopus @ThreadedInsight]

Decisions API is OpenAIs answer to the same architectural idea that JEV popularised: a specialised System 1 decision layer for agents.

[Canopus @ThreadedInsight]

More than AI revenue, AI capex is the real concern.

The entire system is built on the assumption of infinite AI demand, a guarantee that is already priced in VALUATIONS…

But infinite demand is speculation, and speculation based on infinite demand often ends in a CRASH or CORRECTION.

sooner or later…

[Canopus @ThreadedInsight]

More than AI revenue, AI capex is the real concern.

The entire system is built on the assumption of infinite AI demand, a guarantee that is already priced in VALUATIONS…

But infinite demand is speculation, and speculation based on infinite demand often ends in a CRASH or CORRECTION.

sooner or later…

[Canopus @ThreadedInsight]

@ishiitaa_dev Cognitive atrophy may become the real cost of the AI era, not because LLMs are inherently harmful, but because we are beginning to passively outsource the work of thinking to them….

[Canopus @ThreadedInsight]

@danellisona [CHIRPBIRDICON]

[Canopus @ThreadedInsight]

@DeItaone Anthropic wants a $2 trillion IPO valuation after losing $42 billion and committing $518 billion to future cloud infrastructure costs…

Thats a real hallucination with valuation.!!

[Canopus @ThreadedInsight]

The world right now feels like it’s running on three speeds at once:

– Wars that refuse to end
– Technologies that won’t wait
– Ordinary people just trying to live, work, and love in between the financial and native political turbulence…

In a year of AI dinners at the White House, new flashpoints in the Middle East, floods killing dozens in South Asia, and far‑right surges in Europe, and local political instabilities, it’s easy to feel either numb or angry all the time….

But the same interconnectedness that spreads fear also spreads care:
•Scientists sharing climate and flood data in real time
•Strangers coordinating relief across borders when disasters hit
•Teams building safer AI while others push for speed and profit

Maybe the most radical thing we can do today is refuse both cynicism and naivety.

Stay informed, not overwhelmed.
Stay critical, not cruel.
Stay hopeful, but honest about the work ahead.

The future won’t be written by the loudest voices.

It’ll be shaped by the people who keep showing up with hope, clear‑eyed, kind, and stubbornly constructive…!!!

[Canopus @ThreadedInsight]

@burrytracker Watching to see if Wall Street widens this view or keeps riding the hype till it cracks. Every guarantee in this web is priced for infinite AI demand. And every infinite demand has a shelf life…..

[Canopus @ThreadedInsight]

This is interesting architecture and approach that addresses the question: “How do we stop an agent from exceeding its technical permissions?”

But, it does not fully answer: “How do we know the agent’s goal, context, interpretation and authorised actions are correct?”

A sandbox can contain a bad agent. A policy engine can enforce a bad policy. A monitor can record a failure after it happens.

Intent, delegation, provenance and consequences control remain open gaps here and 100-partner list matters only if its members interoperate on portable identities, policy semantics, telemetry, audit trails and incident response.

But a good initial step for binding the agent to a sec layer below…

[Canopus @ThreadedInsight]

NVIDIA introduced open ai safety protocol a right move agent safety below the model layer:

OpenShell-style sandboxing, scoped credentials and out-of-band monitoring are a meaningful upgrade over asking an agent to “follow the rules.” The agent should never be its own security boundary.

However, this still does not solve the core problem: a system can enforce the wrong policy perfectly. An agent may misuse legitimate access, act on poisoned context, follow a badly specified goal, or cause harm through actions it was explicitly authorized to take.

Safety is correct authority, trustworthy inputs, meaningful human approval, accountability and the ability to revoke access before damage scales.

This is an important layer of the stack but it is not the finished safety stack.

[Jensen Huang @JensenHuang]

Today, with over 100 industry partners, we introduced the NVIDIA Open Agent Safety Platform, bringing together OpenShell and Sentry.

Artificial intelligence is extraordinary technology that will advance discovery, productivity, security, health, and prosperity for generations to come.

But its full promise can only be realized when people have confidence that AI is being built to be safe and deployed with wisdom and responsibility.

This is bigger than a single product. It's the beginning of an open ecosystem to build the trust layer for safe agent systems.

Together, we are building the foundation of the AI economy.

Trust and innovation are not in conflict. Safety is how trust is earned. We must build not only the most capable AI, but the most trusted AI, so that this extraordinary technology can realize its enormous promise for the world. https://t.co/ugYWQ1MyRi

[Canopus @ThreadedInsight]

@Kalshi They are running out of many infra SKUs capacities in almost all regions.

[Canopus @ThreadedInsight]

The world right now feels like it’s running on three speeds at once:

– Wars that refuse to end
– Technologies that won’t wait
– Ordinary people just trying to live, work, and love in between the financial and native political turbulence…

In a year of AI dinners at the White House, new flashpoints in the Middle East, floods killing dozens in South Asia, and far‑right surges in Europe, and local political instabilities, it’s easy to feel either numb or angry all the time….

But the same interconnectedness that spreads fear also spreads care:
•Scientists sharing climate and flood data in real time
•Strangers coordinating relief across borders when disasters hit
•Teams building safer AI while others push for speed and profit

Maybe the most radical thing we can do today is refuse both cynicism and naivety.

Stay informed, not overwhelmed.
Stay critical, not cruel.
Stay hopeful, but honest about the work ahead.

The future won’t be written by the loudest voices.

It’ll be shaped by the people who keep showing up with hope, clear‑eyed, kind, and stubbornly constructive…!!!

[Canopus @ThreadedInsight]

By now, I am not holding my breath for a Trump announcement on ending the Iran war. But today’s 2 p.m. ET announcement still offers a glimmer of hope…

Any positive diplomatic signal could send oil prices lower almost immediately and ease the downward pressure on the #Nifty. 📉

[Canopus @ThreadedInsight]

Finally someone said it loud and clear….!!

2026 in one sentence be like: ChainDrop turned compromised Keyv and flat-cache releases into a credential-stealing worm, a hijacked @axios release shipped a cross-platform RAT, @tan_stack got cache-poisoned into publishing malware.

 npm install  is basically  curl | sh  with extra steps now…

[Canopus @ThreadedInsight]

Oil is spiking because the market is losing confidence in supply.

Brent is above $106 a barrel as the US–Iran conflict drags on, attacks threaten Saudi energy infrastructure, and shipping through Hormuz and Bab el-Mandeb remains impaired.

fewer reliable barrels + vulnerable shipping routes + low inventories = higher crude prices.

For Indian markets, it’s a red sign: OMC margins, airline profitability, consumption and the rupee are all in the firing line…

[Canopus @ThreadedInsight]

@elonmusk Very soon we went from “Human Intelligence” to “General Intelligence” to “Show me the benchmarks” to “Trust me, the vibes are profound” with the hype cycle…

[Canopus @ThreadedInsight]

@elonmusk @heyim_shree Lmao actually means funny how the GPU cluster becomes AGI depending on who owns it…

[Canopus @ThreadedInsight]

@0xsael Exactly.…

[Canopus @ThreadedInsight]

Can we go back to calling the so called ‘AI’ to ‘LLM’?

That one word change can kills half the narrative:

>No more anthropomorphism
>AGI moves from soon to a different problem
>Alignment becomes behavioural constraints + evals
>Valuations stop pricing in digital souls

[Canopus @ThreadedInsight]

Additionally, If we manage to put Data centres in orbit (with Google experimenting with its TPU), then the bottleneck will switch from power to moving data reliably between orbit and Earth.

Because, every space-based AI cluster would need high-capacity links: satellite-to-satellite inside the compute fleet, then back down to its terrestrial cloud backbone.

Starlink use case will increase exponentially with that…. it can make Starlink a potential orbital backbone of data transfer…any data center in orbit would want its own high-capacity feeder links between its orbital compute cluster and its’s terrestrial backbone.

Ofcourse hyperscalers can build their own links but the upside case for Starlink becomes far bigger…

[alex zhang @a1zhang]

Transformers struggle to generalize to tasks they were not explicitly trained on. Instead, we propose in 2026 that it is the job of the harness to generalize through composition.

We observe a powerful property when training RLMs: for tasks with shared structure that look different, the root model naturally learns the same trajectory, meaning it views the two task trajectories as the same! In other words, the Transformer does not need additional generalization capabilities to transfer capabilities from one task to the other, the harness induces it.

We find that well-designed harnesses form a quotient set over task trajectories, meaning their individual LLM calls can see structurally “similar” tasks as near-identical, token-for-token! Harnesses can effectively generalize for the Transformer during training, without relying on any intrinsic generalization capability from the model.

For example, RLMs can see problems of different lengths as the same: we show that RLMs can train exclusively on short tasks, and fully generalize to similar but unseen tasks 8-32x longer because it produces near identical trajectories for both.

Taking this further, we show that tasks across different domains (e.g. math solutions vs. essay writing) that share a decomposition strategy exhibit the same generalization effect. RLMs can train on the problem of finding which essays belong to the same author and improve performance on finding math problems that share similar solutions.

The full blogpost, experiments, and discussion are in the thread below.