AI Factories. Balanced Accelerationist. WEKA CAIO, CNCF kubernetes founding board, Post-PKI.
Seat 14D
Joined July 2008
- Tweets231K
- Following8.1K
- Followers8.8K
- Likes322K
Pinned Tweet
Black October is the latest reminder playing chess is a better cyber strategy than checkers
Is there any doubt left that cybersecurity is rapidly transforming into a battle of Red Agent Swarms attacking infrastructure that can only be defended by Blue Agent Swarms, also at machine speed?
Meaning cyber is now a war of token attrition. Losers can't convert energy to tokens at the highest speed with lowest total cost, while winners can.
b/acc, context platform engineer retweeted
A new post from @zartbotF.
It’s an essay on DeepSeek open-sourcing its much-discussed “money-printing machine.” Zartbot argues that DeepSeek’s unique organizational setup is what enabled it to achieve an 80% gross margin.
At other labs, for example, each team has its own KPIs, leaving little room for compromise.
Even if an infrastructure engineer proposes compressing the KV cache further, they’ll inevitably face pushback from the algorithm team if it hurts model performance.
But the most interesting point in the piece goes a step further: even if the algorithm team agrees to experiment with a new architecture that enables greater KV cache compression, the experiment may fail to produce a leading model.
We saw this with V4. When DeepSeek V4 Pro came out and lagged behind GLM 5.3 and Kimi K3, everyone started saying DeepSeek was past its prime.
But Zartbot argues that this was one step back to take two steps forward, and that the work played a major role in the subsequent evolution of 4.1 Flash with CSA2.
A short but fascinating read. I’d encourage everyone to check it out.
mp.weixin.qq.com/s/qlTlTiWoK…
b/acc, context platform engineer retweeted
We're open-sourcing our OrcaCyber DeepSeek V4.1 Flash 1017 (Abliterated GGUF) weights — one of the strongest uncensored/abliterated DeepSeek V4.1 weights we've tested, with virtually no capability loss.
- 91.5% fewer refusals
- 92% MMLU vs. 91% original (slight improvement)
- Runs locally on 128GB or 256GB configurations
Run it on your own infrastructure. No cloud required.
Cyber Tier API: orcarouter.ai/models/orca/or…
Huggingface: huggingface.co/orcarouter/De…
b/acc, context platform engineer retweeted
Everyone who owns hardware with abominable software (like Sonos players) owes a debt of gratitude to Venkatraman Srinivasan and the rest of the eMudhra gang in Bengaluru, Karnataka. Without their gross negligence, it'd be a lot harder to counterfeit certificates and fix the sw🙏
b/acc, context platform engineer retweeted
Hot take from S.F. compute:
“Within the next 6 months, most AI startups will not be able to buy compute. The world’s compute is sold out. To get capacity, you’ll need to sign up for a new cluster. However, it’s becoming increasingly difficult to get capacity online from folks who are not investment grade.
Startups are not investment grade.”
I hope this is wrong, but I think many entities are unprepared for compute shortages and demand being much higher than what Wall Street expects, and so high demand and high interest rates will cause a big problem for everyone.
b/acc, context platform engineer retweeted
Models are are now approaching a massive general jump in cyber capabilities. From the Kimi K3.1 launch post:
'During development, we observed vulnerability research performance improving steadily as post-training scaled. Tracing the source of this trend, we found that cyber-related content was embedded within the general mid-training corpus, and cyber-related tasks were present in the RL environment mix — a natural byproduct of training on broad data. In preliminary experiments on internal models, we found that without data removal at either the mid-training or the RL stage, downstream cyber capability would increase substantially.'
Moonshot is intentionally throttling cyber capabilities. From the blog;
'First, we aim not to release open weights that would extend the absolute frontier of cyber capability. K3.1 Pro's offensive cyber capability is, by design, below what our pipeline can produce, and remains behind that of frontier proprietary models. Second, we intend to progressively raise the cyber capability available in open weights. We believe making K3.1 Pro available in a non-discriminatory way benefits people who need dual-use capabilities for defensive purposes.'
b/acc, context platform engineer retweeted
SITUATION EXPLAINED: China's Kimi K3.1 Pro got so good at hacking that Moonshot stripped its cyber skills before open-sourcing it.
• Kimi K3.1 Pro has 2.8T parameters (104B active) and a 1M-token context window, with full weights due by Nov 4
• It nearly matches Opus 5.5 on Terminal-Bench 4.0 (65.7% vs 66.4%) and beats it on SWE Marathon, HLE with tools, and OfficeQA Pro v2
• API pricing is $2/$10 per million tokens, a third cheaper than Kimi K3
• Removing cyber data from training dropped its score on real exploit tasks from 90.5% to 81%, now below DeepSeek V4.1 Flash at 85.7%
• Moonshot pledged not to release open weights that "extend the absolute frontier of cyber capability"
• It still plans to raise the cyber capability of its open weights over time, saying defenders need dual-use tools too
@theojaffee: "I don’t think I’ve ever heard of a Chinese model or a Chinese lab nerfing any of their models on anything ever."
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.
🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal.
🔹 Full-project coding: games in Unreal, Unity & Godot.
🔹 Excels at parametric 3D modeling in Blender & CAD.
🔹 More reliable at finance work: Accurate modeling with hallucination down by 65%.
🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.
Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Open weights by Nov 4, 2026.
🔗 API: platform.kimi.ai
🔗 Tech blog: kimi.ai/blog/kimi-k3-1-pro
Breaking AI news
I’ve left Anthropic after more than a year of working on Claude Code and seeing it scale larger than I ever imagined.
I have really loved working at Ant and believe deeply in the mission, and most of all, the idea of a race to the top.
The idea that if we work hard and responsibly, we can raise the bar and inspire others to great outcomes. And if we all push each other to be better, things can be radically good. I will carry the spirit of this with me forever.
My colleagues at Anthropic are the most talented, thoughtful, competent, and earnest people I have ever met. I am glad they are at the wheel and trust them deeply.
To the community on X, thank you for everything. You’ve made this the most fun (and sometimes stressful) job I’ve ever had.
For me, it is time for a bit of a break before diving into something new.
If you want more excellent Claude Code content, you should follow @addyosmani , @lydiahallie, @_catwu, @ClaudeDevs and of course @bcherny who started it all.
b/acc, context platform engineer retweeted
It’s insane the extent to which prompts are shareable programs now.
A very clear prompt can provide enough structure for an agent to deterministically recreate software.
The latest models can pack a whole lot of software in very few words.
b/acc, context platform engineer retweeted
Many recent lone-actor incidents used ARTEX with DeepSeek 4.1 Flash as the primary model (including the South Korean bank attacks documented by CrowdStrike). This is an abliterated version of that model with the guardrails removed. Great for defense too, but as always, dual-use.
Now live: abliterated-model-large-flash (our version of DeepSeek-V4.1-Flash-1017). It is a cheaper alternative to our GLM-5.3-based model for product owners and defenders who need access to the most cyber-capable models.
- 1M context
- Vision
- Zero input/output retention or training
We believe paying customers should be in control, and over the past few months, leading experts and major brands have adopted our models.
abliteration.ai
b/acc, context platform engineer retweeted
Now live: abliterated-model-large-flash (our version of DeepSeek-V4.1-Flash-1017). It is a cheaper alternative to our GLM-5.3-based model for product owners and defenders who need access to the most cyber-capable models.
- 1M context
- Vision
- Zero input/output retention or training
We believe paying customers should be in control, and over the past few months, leading experts and major brands have adopted our models.
abliteration.ai
b/acc, context platform engineer retweeted
i got curious about what dario (anthropic CEO) is predicting next, so i went through all his recent essays and interviews
if he keeps being correct, here's what our future timeline looks like:
> 2026: someone builds a $1B company with just 1 employee
> late 2026-early 2027: AI matches or beats nobel prize winners in most fields, from biology to math
> 2027: chinese AI catches up to Mythos, the model anthropic kept from the public because it's so good at hacking
> 2027-2028: AI gets better than every human at almost everything, with millions of copies running 10-100x faster than us
> 2027-2028: AI starts building the next generation of AI on its own
> 2028: AI is using 100 gigawatts of power, about a fifth of all the electricity america uses
> by 2030: AI disrupts half of all entry-level office jobs
> before 2030: AI brings in trillions of dollars a year, and the economy grows 10-20% a year
> early 2030s: cancer deaths drop 95%, alzheimer's becomes preventable, and nearly every infectious disease can be prevented or treated
> early 2030s: most mental illness, like depression and PTSD, becomes curable
> 2030s: sub-saharan africa gets as rich per person as china is today
> 2030s: the human lifespan heads toward 150
b/acc, context platform engineer retweeted
the PHP vulnerability story is actually insane lol Im a bit scared
we’re talking about a memory corruption bug that may have existed for almost 10 years.. yea .. buried deep in PHP’s memory management.
the issue involves heap manipulation…where memory can be corrupted in ways that potentially allow remote code execution.
for years vulnerabilities like this were incredibly difficult to investigate. understanding the memory behavior alone required serious expertise, time and resources.
fast forward to 2026 and researchers are using AI to investigate these same complex problems.
and this is what blows my mind.
we’ve spent decades building on top of legacy systems and now AI is making it easier to understand what has been hiding underneath them + PHP still powers a massive portion of the internet.
the barrier to discovering vulnerabilities is getting lower for security researchers AND attackers.. ugh .. i think we’re about to see a very interesting era of cybersecurity.
and maybe it’s time we rethink how we maintain the infrastructure running half the internet lol
b/acc, context platform engineer retweeted
yeesh, that's quite a bit above my forecast
poor 1017 gets bodied (at least it mostly beats GLM 5.3)
Moonshot AI claims to stand on the public frontier.
if it's token-efficient, they have secured their position as one of the live players
we shall see
b/acc, context platform engineer retweeted
I see this chart a bit differently:
- New companies are able to design their org/GTM from zero, and are adopting AI furiously
- Enterprises have cultural, political and technical reasons to move much slower on AI adoption
That’s pretty much it
b/acc, context platform engineer retweeted
500T tokens. That’s how much traffic @sonicaghi says is moving through @Kong’s gateways today.
Run the math on 140x YoY growth and the volume roughly doubles every 7 weeks. At that rate, a quadrillion is a near-term milestone.
Cheaper tokens won’t cap that. They’ll spread it across more models. Enterprises asked Kong how to secure, govern, and cost-control API calls a decade ago. Now they’re asking about tokens.
His read: “In a non-deterministic software enterprise world, the need for deterministic AI gateways will just increase.”
Kong CEO @sonicaghi reveals token traffic through its gateways is growing 140x year over year, approaching 1 quadrillion tokens:
"The number of tokens is growing exponentially. For the subset that we see, we are running about 500 trillion tokens, going to a quadrillion soon, because it’s growing 140x year over year."
"The cost per token is going down, and inference will still be exploding. We’re gonna live in this enterprise world that is more multi-model, cheaper tokens, cheaper levels of intelligence."
"The next pieces are, how do we secure it? How do we govern it? How do we cost control it?"
@kong
b/acc, context platform engineer retweeted
Doesn't this really mean that all the Internet infrastructure and e-commerce companies making tens of billions of dollars finally have to start spending some money on tokens to patch basically every single library they depend on in their entire stack?
Big Tech has been taking a free ride on these libraries, and suddenly it stopped being so free. Now it's a bit of a liability, so it's time to contribute something. That something should be billions of dollars in tokens to patch everything and make it super secure.
b/acc, context platform engineer retweeted
This one maybe as bad or worse than the PHP ones. Get used to lots of patching.
