The other thing that kind of bugs me (pun intended), is the "31% of intrusions are the result of software exploitation" statistic.
It hides two different problems:
Zero-day exploitation - the bad guys find a vulnerability before the good guys, then use it to hack into systems (a problem finding and reporting vulnerabilities reduces).
N-Day exploitation - someone reports a vulnerability, a fix gets published, but then the threat actors rush to use the information from the fix to build an exploit before downstream users get a chance to install the fix (a problem finding and reporting vulnerabilities exacerbates).
Since organizations typically take a long time to install security fixes, N-day exploits are extremely powerful, and also make up the majority of all software exploitation. There's also 100x more people who can create N-days from known vulnerabilities than can find and exploit Zero-days.
The more vulnerabilities found, the more vulnerabilities N-day attackers have to choose from, and the more software fixes that downstream users have to juggle.
Orgs struggle to even install the monthly Windows update within a month of it going live. Now mix in "every single product you use has critical vulnerability fixes every patch cycle so you have to install all 4,000 patches every month". And yeah.... It's pretty bleak.
That's not to say vulnerability research shouldn't be done, isn't important, or isn't a net long term benefit (I feel like I have to say this because half the people on this fucking website have zero capacity for nuance and try to read every post in the most disingenuous way possible).
But since the majority of exploitation is N-day exploitation, and vulnerability research increases N-day exploitation, using exploitation rate as a justification for pouring more resources into find vulnerabilities (but not helping fix or patch them), is just circular reasoning.
Plenty of good arguments for why vulnerability research is important, but "we're showing the attackers more vulnerabilities that they can use, which is leading to them using more vulnerabilities" is not one of them.
Equal effort needs to go into all part of the vulnerability management pipeline for this to work. But currently the effort drops by a factor of 10 from finding bugs to fixing bugs, then another factor of 10 from fixing bugs to ensuring the fixes get installed.
Just reducing the disclosure window and telling everyone "don't worry, I'm sure you'll figure it out" is extremely likely to make the problem worse, not better.
Will DePue (@willdepue), a former OpenAI employee who worked on the video model Sora, says he can tell which chip was used to make an AI model just by looking at the model. Four of China's top AI labs used a chip that the US won't sell to China for their most recent models, according to DePue.
Before a model begins training, it gets filled with billions of random starting numbers. Those numbers are generated from a "seed": the same seed will produce the same numbers. The chip that does this divides the work among 2,048 workers for every processing unit it has. Nvidia's older H100 chip has 132 processing units; its newest Blackwell chips have 148 or 152.
If you use the same seed on different chips, the same numbers get put in different places because the number of processing units is different. The final layout ends up encoding the number of workers that distributed the numbers, and thus the chip that did the work.
Training then changes most of the starting numbers, but not all of them. Some are never touched and stay at their starting values.
Using a Blackwell chip, DePue reproduced the untouched starting numbers, in the exact order, from DeepSeek's V4 Flash model, released in April. He then found the same Blackwell fingerprint in six models from DeepSeek, Moonshot, MiniMax, and Ant Group.
The discovery is notable because limiting China's access to the chips is the cornerstone of America's efforts to slow China's AI advances. Since 2022, the US has banned sales of its most powerful AI chips to China, arguing they would make the country's military planning faster and more accurate.
The goal was to keep Chinese labs a generation behind. The Blackwell chip is more than twice as powerful as Nvidia's previous flagship product, and all four labs released their first models built on Blackwell within six months of each other.
By DePue's estimate, the final training runs alone used around 12,000 of the processors, a stack of hardware worth more than half a billion dollars and illegal to ship to China without a US export license.
But as DePue points out, Blackwell chips can be rented from data centers outside China: what he has shown is access to the hardware, not that it was illegally sent to China.
US export rules regulate where a chip is shipped, not who uses it. If a Chinese lab rents time on a Blackwell located in Thailand, that's not clearly a violation, and a proposed regulation to plug this loophole has been stuck in draft since the summer.
It's also possible that the trick DePue used to detect the chips won't work a second time. He couldn't find his Blackwell signal in the model that DeepSeek released after that, V4.1 Flash, which he attributes to the company changing how it sets its starting numbers. Washington's main tool for holding back Chinese AI development didn't stop any of the top four labs from training on the latest generation of chips, and the one technique for detecting that fact is something the labs can switch off.
Why is Canada getting poorer than US. Not because it has less, it is because it cannot churn out more.
Credit for the video to Neighbors Apart
Grab a coffee: 9min.
Even be disrupters are being disrupted! Everything is happening litening fast. Will there be anything left to disrupt? World would need a new economic model pretty soon it seems!
Great ro see Indian AI company building state of the art models. There are a lot of innovation happening below the surface in Indian AI scene. Scale n consistency is the new frontier for India
An Indian AI startup just achieved SOTA with their music model.
Taan v1-preview is the world's most controllable foundation music model ever. It's 3x better with instructions than Suno.
As someone who has spent HOURS playing with Suno, rolling and re-rolling the same prompt to try and get the sound I'm after (to no avail - I usually just end up settling with whatever Suno spits out, even if it's not what I wanted), it's crazy to me that you can tell Taan "start the violin at exactly the 11th second" or "first 8s tense, pause for 2s, and then the beat drops at 10s" and it will execute your instructions to a T.
This is the way generative music models always should have been.
Oh, and it's trained on 100% licensed music + you get perpetual rights to use the music you generate commercially.
🚀 گوگل یک PDF نهصفحهای درباره Harness Engineering منتشر کرده؛ مفهومی که به نظرم باید هر کسی که با AI Agent کار میکند، جدی بگیرد.
فرمول اصلی ساده است:
Agent = Model + Harness
یعنی مدل هوش مصنوعی بهتنهایی ایجنت نیست. چیزی که آن را به یک نیروی اجرایی قابلاعتماد تبدیل میکند، زیرساختی است که دور آن میسازیم.
اما این زیرساخت از چه چیزهایی تشکیل میشود؟
۱. Guides: قوانین و دستورالعملهایی که تجربه شکستهای قبلی را به دانش دائمی تبدیل میکنند.
۲. Sensors: تستها، Linterها و ابزارهای اعتبارسنجی که اجازه نمیدهند هر خروجی اشتباهی به مرحله بعد برود.
۳. Agentic Loop: چرخه برنامهریزی، اجرا، بررسی و اصلاح؛ با محدودیت بودجه و امکان ارجاع به انسان.
۴. External Memory: حافظهای خارج از مدل برای حفظ تصمیمها، وضعیت پروژه و دستاوردهای جلسات قبلی.
۵. Permissions: کنترل دسترسی ایجنت به ابزارها و تعیین مرز استقلال آن.
۶. Observability: ثبت و پایش تمام فعالیتها، هزینهها، خطاها و تلاشهای مجدد.
اما داستان اینجا تمام نمیشود!
چند پروژه متنباز برای ورود عملی به این حوزه:
🔹 Google Harness²
https://t.co/DrQfFijj6q
🔹 Learn Harness Engineering
https://t.co/Z4OF17F6Ka
🔹 Awesome Agent Harness
https://t.co/hI6WSGS4vn
🔹 Harness Engineering
https://t.co/86n27omyXE
و حالا یک مفهوم جذابتر:
Graph Engineering
تصور کن چند ایجنت مستقل داشته باشی که هرکدام Harness مخصوص خودشان را دارند.
یکی تحقیق میکند، دیگری تحلیل میکند، سومی کدنویسی میکند و چهارمی خروجی را اعتبارسنجی میکند.
حالا اگر یکی شکست بخورد، شکستش میتواند به سیگنالی برای اصلاح عملکرد بقیه تبدیل شود.
اینجا دیگر با یک چتبات یا ایجنت ساده طرف نیستیم؛ با یک سیستم اجرایی چندایجنتی روبهرو هستیم.
یک نکته که نباید فراموش کنیم:
Hermes را میتوان بهعنوان محیط اجرای ایجنت در نظر گرفت؛ جایی که ابزارها، مهارتها، حافظه و گردشکارها کنار هم قرار میگیرند.
Harness Engineering کمک میکند این محیط را قابلکنترل، قابلاندازهگیری و قابلاعتماد کنیم.
به نظرم آینده متعلق به کسانی نیست که صرفاً مدل قویتری انتخاب میکنند؛ بلکه متعلق به کسانی است که سیستم بهتری دور مدل میسازند.
مدل، مغز است؛ Harness، سیستم اجرایی؛ و Graph، شبکه همکاری مغزهاست.
This is insane. You can literally type a sentence and watch a Transformer process it.
This interactive explainer runs a GPT-style model in the browser and lets you follow the whole path:
tokenization → embeddings → Q/K/V → attention → MLP → logits → next-token probabilities
You can inspect attention maps, intermediate values, and change temperature, top-k and top-p to see how generation changes.
If Transformers still feel like boxes connected by arrows, this is worth opening.
https://t.co/vkyxOqDVll
Wow! Lot to learn from this. Lotd of interesting technology and tools to build an AI agent using Claude Fable. Interesting way to extract features from the existing bots by popular labs
I stole the best part of every AI agent.
One prompt in my software factory. Fable 5.2 built it in 3 hours:
- Grok Bot's interface and voice
- OpenAI Dots' computer
- Hermes as the harness underneath, any model
- Muse's editable soul and memory
- iMessage from Instinct
I sent one message in T3 Code and came back to a live bot platform that picked up when I called it.
Full 36 min build, whole prompt on screen. Open source.
Watch.
Claude Fable 5.2 is simply the greatest AI model ever
• By far most intelligent
• Fast and efficient
• Understands your intentions on a level I've never seen from a model before
In this video I cover why this model is so good and a workflow that will 10x your productivity:
Google published a paper that might end the PhD.
It’s called “ScientistTwo”
For centuries, the bottleneck to human knowledge has been the human brain.
It takes five years of grueling labor to read the literature, find a gap, formulate a hypothesis, run the experiments, and write the paper.
But, Google automated the entire pipeline.
They built an AI that reads a problem, forms hypotheses, runs experiments, writes the paper, AND simulates its own peer review. no human touches it.
It reads millions of papers to find blind spots in the current science.
It generates novel hypotheses.
It designs the experiments. It writes the code to test them.
When the code fails, it debugs itself.
When the results come in, it analyzes the data, generates the charts, and writes the final academic paper.
It even simulates a brutal peer-review process to critique and improve its own work before it considers the project finished.
This isn't an LLM summarizing a PDF. This is a fully autonomous digital scientist.
What takes a human candidate half a decade of massive intellectual effort... the AI accomplishes in a fraction of the time, at a fraction of the cost.
The implications are staggering.
If you can spin up 10,000 autonomous researchers overnight, the rate of scientific discovery is about to go vertical. We are moving from artisanal, human-speed research to industrialized knowledge creation.
But it raises an uncomfortable reality for academia.
When an AI can read the literature, design the experiment, execute the code, and write the paper...
What exactly is the human supposed to do?
Google just open-sourced the plugin thousands of its engineers use every day to stop AI from shipping sloppy code.
It's called the Elephant-Goldfish Model.
No code until the plan holds up. No AI grading its own work.
The Elephant is one long-running AI session that remembers everything. It writes no code. It questions you and keeps the record.
The Goldfish are fresh sessions with no memory, so they don't share the Elephant's blind spots.
How it works:
→ egm-brainstorm: you and the Elephant debate options, with no code allowed → egm-draft: you write a Markdown design doc with the plan and the exact files to change → egm-review: a Goldfish explains the design back. If it gets it wrong, the doc gets redrafted → More Goldfish act as hostile reviewers, hunting for edge cases and gaps in the logic → A final Goldfish makes the go/no-go call
Link: https://t.co/SxtMb7wyNR
India has achieved remarkable progress on the environmental side over the past decade. I don't see global media houses praising PM Narendra Modi enough.. but they should do.
1) COPs target for India was having 50% of its cumulative electric power installed capacity from non-fossil fuel sources (solar, wind, large hydro, and nuclear) by 2030.
India achieved this five years ahead of schedule! By mid-2025, non-fossil capacity crossed 248 GW out of a total installed capacity of over 484 GW.
2) India said it is reducing the emission intensity of its GDP (emissions per unit of GDP) by 45% by 2030. It has already achieved a reduction of 36% well ahead of the deadline, driven by energy efficiency standards, cleaner industrial processes, and grid modernization.
3) Under its pledge to create an additional carbon sink of 2.5 to 3 billion tonnes of CO₂ equivalent through forest and tree cover, India expanded its forest footprint to over 25% of its total geographical area, adding billions of tons in carbon stock.
4) PM Narendra Modi has made a new commitment at COP26, scaling non-fossil energy capacity to 500 GW by 2030, cutting projected carbon emissions by 1 billion tonnes by 2030, and achieving net-zero emissions by 2070.
Amazing contribution to mother Earth by the world's most populous country 🇮🇳
If Prasar Bharati can spend ₹15 crore a year on a DD News show featuring a star anchor, while the ISRO Chairman’s basic pay is about ₹27 lakh a year, shouldn’t we rethink our priorities? If India wants to conquer space and attract our finest scientific minds, our rewards must match our ambitions.