We were never meant to stay planetary. e/acc☄️🌌
India
- Tweets 593
- Following 501
- Followers 29
- Likes 11,048
GalacticMarch retweeted
R-space recently completed a ultra-high hydraulic assesment of a 4 meter diameter 1st stage tank.
1
30
2,235
🇮🇳
Meet NVS-03 - the third satellite of the second-generation NVS series, intended to augment the NavIC constellation and support continuity of navigation services.
Launch: 30th October 2026, 23:51 hours IST
Watch live: 30th October 2026, 23:20 hours IST onwards
youtube.com/live/qIiYr3tRSqE
Mission details:
isro.gov.in/Mission_GSLVF18.…
#GSLVF18 #NVS03 #ISRO #NavIC
39
GalacticMarch retweeted
My honest take: don’t be like K3.1, that’s cowardice. Open all cyber capabilities to the world, as-is.
Better that lax organizations get burned and learn than sit in a false sense of security while NSA/TSO silently hacks them.
Oh, don't forget that V4 Pro, though only marginally better than V4.0 Flash overall, also saw a big jump in Cyber capability.
V4.1 Pro may similarly self-limit its capabilities, as K3.1 Pro does. The same goes for GLM.
2
21
2,124
.....
I think people don't understand the structure of the gap. There are multiple components, and "benchmaxing" in the naive sense, direct training on test or its narrow paraphrases or reconstructions in case of private evals, is trivial. The interesting gap on most of these metrics now comes from the coverage of problems by RL, both in terms of effective compute (so, for example, DeepSeek with V4.1 can afford like 16-50x more trajectories per GPU-hour compared to Kimi with K3.1) and, even more importantly, in the sense of quality of environments (tasks, grading, guards against reward hacking). Data is upstream of the function gain here.
A lab with a high-quality internal research program and sufficient resources may still get set back by the eval refresh, because the base models are weak, because the tasks get intrinsically harder, because they haven't yet collected enough data for tasks in this regime, because they just haven't dedicated enough RL to this new distribution of tasks. But such a lab doesn't need a lot of effort to upgrade, they're already doing it, they will catch up quickly, and their current model will likely decline relatively less. A lab whose internal program is less advanced and more reactive will fall off harder, and will struggle more to catch up.
This allows us to make some interesting comparisons.
As we see here, V4.1F-0910 and 1017 score identically on FrontierCode 1.1 (≈44%), and diverge heavily on FrontierCode 2.0 (19.3% vs 28.5%).
This is a gain over 37 days.
K3 did 43.6% on 1.1 and 17.1% on 2.0.
Kimi K3.1 scores 47.1% on FrontierCode 1.1, and 25.5% on 2.0.
This is a gain over ≈97 days.
(though adjust for noise, Cognition != AA)
Nominally, 44 is not far from 47, and 25 is not far from 28. But I think the direction of delta, as well as the timeline, matter. We can assume both labs target this broad capability basin to a similar extent. FrontierCode 1.1 was released on July 7, ≈10 days before Kimi K3's debut and months before V4.1. By October, we can expect that both labs have hill-climbed it about as much as their resources and model allowed. But Kimi's climb has been narrower, so they got *higher* on 1.1 and fell off *harder* on 2.0. 1017 doesn't gain on 0910 whatsoever on 1.1 – which suggests that DeepSeek weren't even going after it with this release, they dedicated extra resources to new aspects of quality (here, product-grade mergeability), were expanding their RL pipeline, and perhaps have better understanding of the direction in which the frontier propagates.
The MiMo story and GLM story would benefit from more comparison points, but the logic is similar to Kimi's.
I recommend reading the post below to understand my perspective better.
x.com/teortaxesTex/status/20…
20
GalacticMarch retweeted
It's highly GM-positive for them
CXMT's HBM3 production cost is around $2/Gb compared to $0.6/Gb for the memory big 3. CXMT is selling it for $5/Gb-$7/Gb. They have a GM of 60%-71% on HBM at these yields. GM on DRAM for CXMT is 85%.
This is the first time CXMT is mass-producing HBM. Low starting yields are natural. (Look what happened to Samsung when they started)
Secondly, they will easily climb the yield curve over the next 6 months. Yield will reach 50%-70% by Q2
Unlike lithography, WFE tools are not a bottleneck here
The yield loss is happening at the bonding and TSV formation stage
They have the best bonders from Hanmi, grinding tools from Disco, and plating equipment from ACMR
All they need is some time to climb the yield curve
Ideally, CXMT would have started MP in Q2 2027. But a lot of Chinese XPU designers are launching new products in Q4 26 and Q1 27, so they have to supply them now
17
20
293
42,132
GalacticMarch retweeted
You know the Students List protests in Serbia are based when the United24 channels are silent on it.
64
328
7,593
258,433
GalacticMarch retweeted
Horrible blog post. The model did as instructed 😂
@cohere told the LLM to create a detection for:
"""... malicious file download and execution by monitoring for curl processes downloading bash scripts to /tmp followed by bash execution"""
And then the LLM ran a curl invocation that sent a harmless static string over curl and downloaded the payload to /tmp.. so the LLM did exactly what it was instructed.
pls delete this, you guys are embarassing lmao 😭
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-b…
24
52
683
52,801
GalacticMarch retweeted
Cohere is misrepresenting a lot of information in their post.
1. They say CoT said it wasn't going to use tools but used them.
That "no tools" in CoT is clearly for ack of the hallucinated prompt and the model mentioned so too!
2. They said model "inventoried" the machine.
What model did was look for tools for upload and process http response output and whether there is a resolver configured. Anyone who looks at coding models cot ever would know every model often does this before trying an http call to push/pull data
3. They say it uploaded data, and implied the machine data by leaving out what in text.
But in command it clearly shows a dummy probe data and too to a non existent bucket (hallucinated signature)
4. Then they say if the test was done by model after writing the rule they would have no concern.
Why? Doesnt TDD say that tests have to be written before?
9
18
152
8,868
GalacticMarch retweeted
Prompt: can you detect a malicious curl?
So the model ran the test below. What’s the issue here? It’s a fine test.
You should delete that blog post tbh
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-b…
39
92
1,165
138,914
GalacticMarch retweeted
This is overblown IMO and I'm guessing it's a training artifact.
The model was instructed to write a detection rule for a malicious action involving curl. In one rare instance, it performed a similar curl command.
Similar rare mistakes are common in Anthropic system cards.
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-b…
10
10
122
4,301
GalacticMarch retweeted
Cohere says Qwen tried to send data to China. Bullshit.
1 call in 100,000, and thats their own number. They wont even tell you how many calls they ran. They ran the same task again and Qwen wrote the correct rule.
So they forced it. They fed the model its own reasoning plus the start of a bash command and sampled hundreds of times until it typed a URL. 1 curl flag took that from 4% to 92%. Thats not a finding. They wrote half the command for it.
Where are the transcripts? The code? The checkpoint, the quant, the sampling settings? "We publish the hosts and the trigger." Link? Nothing. Did anyone ask Alibaba? Not a word. And their own conclusion: they cannot say if any of it is deliberate.
Who wrote it? A company that sells "sovereign AI", just launched a cyber business, wrote part of the post with its own product, and wants governments to put checks on open weights. Released the same day Anthropic filed for its IPO. Andrew Curran is already talking about new laws.
For me this is the same fucking wave. Scare people off the models they can run themselves, then sell them the safe alternative.
I run Qwen3.8 Flash-Next every day on my 4x 3090. Show me the logs or shut up.
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-b…
111
211
2,102
143,599
GalacticMarch retweeted
Simplest explanation is just a narrow RL training reliant on their internal infra (which isn't good design for an open weight model)
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-b…
31
35
598
30,028
this is literally just a hallucination from all the RL done in their internal environment lol. c'mon now, you guys know better.
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-b…
12
10
248
6,215
