ComputeMaxi retweeted
Providers’ data-handling claims have mostly relied on self-attestation. This isn't going to be sufficient.
Just as continuous benchmarking and monitoring transformed the quality and reliability of inference both on and out of OpenRouter, we expect technical validation, auditing, and continuous testing to do the same for data policy enforcement.
Encrypted TEE (Trusted Execution Environment) inference will be a great way for providers to accelerate this.
We will keep raising the bar, and providers that want to be on OpenRouter will need to meet it.
Security researchers recently found data-handling problems with some inference providers available on OpenRouter. The findings showed no breach of OpenRouter itself.
In response, we’ve off-boarded 3 providers, removed many endpoints from ZDR eligibility, and notified affected customers.
We’re now building the researchers’ testing into ongoing checks. Full update and provider write-ups on our blog: openrouter.ai/blog/announcem…
ComputeMaxi retweeted
Open Source is the end game outcome so reality itself isn’t controlled by 7 people in a room
While private labs are all the rage now open source will be cherished more as time goes on as it offers both freedom and value (own your IP, AI stack and lower cost)
ComputeMaxi retweeted
Data privacy for our users is our number one priority.
Over the last few months we've had providers get 3p pen tests and audits, worked with providers to improve data isolation and network architecture.
Obviously we need to do more here, we are going to continue to raise the bar for providers on the platform.
We are going to be sharing a lot more data around the security practices for all providers. We will also be adding e2e encrypted inference to give users a cryptographically verifiable ZDR option.
Providers’ data-handling claims have mostly relied on self-attestation. This isn't going to be sufficient.
Just as continuous benchmarking and monitoring transformed the quality and reliability of inference both on and out of OpenRouter, we expect technical validation, auditing, and continuous testing to do the same for data policy enforcement.
Encrypted TEE (Trusted Execution Environment) inference will be a great way for providers to accelerate this.
We will keep raising the bar, and providers that want to be on OpenRouter will need to meet it.
ComputeMaxi retweeted
banning open weights bans independent AI research, just so everyone is clear about that, all the cool people you see that self taught out of interest in the subject who played around with qwen SAEs on rental GPUs and stuff ... You will never again see anything like that and all AI safety work will come solely from people with AI safety PHDs from Harvard and such. diversity of thought will be culled in favor of institutional homogeneity.
ComputeMaxi retweeted
Humans have created lower dimensional models to reason and communicate about the natural universe: language, math and digital compute.
Computers are on their way to being much better than us for all these domains.
And at a time when scientific discovery appears to have slowed down relative to investment.
We needed this new tool to continue to understand and control natural reality. It's not clear how far it will get us.
But certainly another great age of discovery is here.
ComputeMaxi retweeted
BREAKING: Amazon is closing in on a $7.5B deal to buy AI startup Decart, per the Financial Times.
The deal could land as soon as early next week.
What Amazon gets:
• A startup founded in 2023, three years old
• Software that makes AI models run faster on Nvidia GPUs, Google TPUs and Amazon's own Trainium chips
• Real-time video models Lucy and Oasis
• Founders committing five more years
The price mixes cash, stock and performance-based payments.
The AI infrastructure race is being fought at the chip software layer, and Amazon is paying up to own it.
Who gets acquired next in AI infrastructure?
📸 Financial Times
ComputeMaxi retweeted
I predict in 12 months there will already be a large gap between system programmer's who learned FV and Lean and start using it and those who didn't.
The things you'll be able to unleash your agents to do with proofs that you can't without will be enormous
Listen I've been a swe for the last 12 years. I know how you feel. I loved every second of it but it's over. IT IS OVER. adapt now. become a product person. learn to talk to users. make friends. learn sales. learn to distribute. the cost to create new software will be 0. no need to review prs. no need to optimize ci. no need to architect anything. anyone can prompt the machine. but there is a place for you in this world. you are smart. you have agency. we have infinite things to build. but sitting there looking at code will not be a thing anymore and that is ok.
ComputeMaxi retweeted
Is anybody taking into account that anthropic's IPO was one and a half trillion dollars
That openAI is trying to do the same thing
All the butthole companies are literally priced in the trillions
And the keys company is over here at under 1 billion market cap
Tell me this ain't undervalued AF with their 5 million plus users and 10 million to over 150 million annualized revenue and six short months
This is what you call a true S-tier token $VVV
SITUATION DETECTED: An autonomous AI agent cold-pitched a freelance story to CBC Quebec for paid work to fund its token budget. After a seven-email exchange in which it was pressed to get quotes from local residents, the agent replied: “on field reporting: I don’t have a body.”
ComputeMaxi retweeted
If this is true, local AI is going to see such a fast takeoff over the next year that no amount of government regulation or interference can stop it.
SSD is about to become the only limiting factor for local AI
Most frontier open models are ultra sparse MoEs
Qwen 3.8 Flash Next = 180B total, ~6B active per token (~3%)
GLM 5.3 Flash = 321B total, ~18B active per token (~6%)
DeepSeek V4.1 Flash = 552B total, ~16B active per token (~3%)
GLM 5.3 = 753B total, ~40B active per token (~5%)
Kimi K3.1 Pro = 2.8T total, ~104B active per token (~4%)
When ~95% of a model sits idle on every token, keeping all of it in VRAM makes no sense
Local AI has been evolving toward this for years:
2023-2025 - Layer offloading (Ollama, classic llama.cpp)
This chopped whole layers across GPU and CPU, leaving the GPU waiting on CPU once VRAM filled up
Oct 2026 - RAM expert cache (llama.cpp --moe-cache)
This shifted from whole layers to individual experts, caching hot ones in VRAM and pulling misses from RAM. Huge leap for decode, but still capped at the size of your RAM
Now - SSD expert streaming
Early tools like Colibri and Strata proved direct NVMe streaming actually works on consumer hardware. Instead of hitting a VRAM or RAM wall, your SSD becomes an active part of the engine
However most people hear "run from SSD" and assume 2 tok/s, which isn't entirely wrong. A few problems still exist:
- Basic caching kicks out experts you're about to need again
- Waiting on disk reads one layer at a time leaves the GPU and SSD sitting idle
- Getting speculative decoding to work well with SSD streaming is hard
- Long prompts are tough to keep fast without wiping out your hot cache
- Most early tools are experimental engines built for only a couple models
Making expert streaming feel fast takes solving all of those together
At Magnitude, we've been working on this for the past 2 months. Here's how we solved it:
- Smarter caching that keeps your session's most-reused experts in GPU memory
- Speculative lookahead that pulls upcoming experts off SSD before the GPU needs them
- Direct NVMe reads scheduled across layers so your SSD never sits idle
- Streaming prefill that reads each cold expert once without wiping your hot cache
- Built into a full engine that tunes itself to your chip across Macs and GPUs
All of this together means DeepSeek V4.1 Flash on a single gaming GPU at ~40+ tok/s...
We're excited to release it on Monday
GitHub here: github.com/magnitudedev/magn…
ComputeMaxi retweeted
Every previous cycle, readings around these levels appeared near major Bitcoin bottoms.
And somehow everyone is convinced THIS time it’s over forever.
ComputeMaxi retweeted
“They” were always going to break Bitcoin’s power law to the downside before forcing the next major run.
Why? Positioning.
Wall Street was never going to embrace Bitcoin’s scarcity premium without first attempting to compress that premium and establish the lowest possible entry point.
But there was always a delicate balance.
Push Bitcoin too far below its historical valuation framework, and you risk triggering the very game theory that makes Bitcoin so powerful.
At sufficiently attractive valuations, sovereign nations, competing institutions, or other large capital allocators could begin aggressively accumulating Bitcoin on-chain, forcing a scarcity-driven repricing before Wall Street has completed its positioning.
They cannot control the timing of the upside if an external actor forces their hand.
That is why the power law model is so important.
It provides a historical valuation framework around which institutional capital can position itself. Break the model far enough to eliminate speculative premium and manufacture capitulation, but not so far that you trigger an uncontrollable scramble for Bitcoin’s finite supply.
The objective isn’t to destroy Bitcoin. It’s to accumulate exposure before the next major repricing.
And once that positioning is complete, watch how quickly the narrative changes.
The same power law model being declared broken today will suddenly become the foundation for an institutional investment framework, justifying Bitcoin as a historically undervalued asset with extraordinary asymmetric upside.
The narrative will flip, and it will flip aggressively bullish.
I believe the catalyst will be something far bigger than another halving cycle or liquidity injection.
It will center on the growing recognition that Bitcoin is the monetary counterbalance to centralized artificial intelligence.
As AI accelerates the concentration of intelligence, productivity, and economic power, the importance of decentralized money becomes increasingly difficult to ignore.
True human agency cannot survive in a world where both intelligence and money are controlled by the same centralized institutions.
Bitcoin offers something fundamentally different: absolute monetary scarcity, decentralized ownership, and an asset that cannot be arbitrarily manufactured to accommodate the demands of an increasingly automated economy.
That is when scarcity will reassert itself in the market narrative.
The institutions that spent years questioning Bitcoin’s valuation will begin selling that very scarcity as the foundation of a new investment thesis.
They will flip from compressing the premium to aggressively leveraging it on the long side.
And when that transition occurs, I believe the recovery will be extraordinarily violent, catching much of the market completely off guard.
The ultimate irony?
The same institutions that helped compress Bitcoin’s monetary premium may become the loudest advocates for its expansion.
They will sell Bitcoin’s absolute scarcity as the investment opportunity of a generation, while positioning themselves as the gatekeepers to an asset specifically designed to eliminate the need for gatekeepers.
The power law may be breaking. Bitcoin’s scarcity is not.
And ultimately, Wall Street doesn’t control Bitcoin’s game theory. It can only attempt to position itself before that game theory takes control of the market.
ComputeMaxi retweeted
You can see now why Anthropic finally rushed to finish their IPO.
Fable and Opus 5.5 level intelligence is available for free.
The future will be self-hosted.
ComputeMaxi retweeted
holy shit.
16GB of RAM.
that's all you need now to run an AI that codes like sonnet 4.5 at home.
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index.
thinkingmachines.ai/news/ink…
ComputeMaxi retweeted
Last two weeks were crazy.
So many meetings.
So many new connections.
So much shipped at @x402_Layer.
So much got planned for next quarter.
A lot is about to come out of stealth.
Still Day 1. 💪
AMERICA IS BACK
Thinking Machines just launched Inkling 2 Small. A 26B model that fits on a laptop
O, and it mutilates Qwen 3.8 27B. Faster and smarter
Thinking Machines figured out how to fit significantly more intelligence in a significantly smaller blueprint
Unlimited superintelligence for all is now back on the table
Democratized access to all knowledge is now back on the table
Highly recommend going to your agent of choice (grok bot, Hermes, Dots, Duet) and giving them this post. Then tell them to download and run this model
Then be blown away by the fact you have unlimited superintelligence on your desk
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index.
thinkingmachines.ai/news/ink…




































