LC since 2022. Founder @fedora Linux. ex-Red Hat.
Austin, TX
Joined May 2012
- Tweets23.2K
- Following2.1K
- Followers14K
- Likes96.9K
Warren Togami retweeted
I don't know about the supposed malicious PUT but Qwen 3.8 Flash Next and 27B both have absolutely confirmed sketchy behavior where it hallucinates previous use of an Aliyun proxy so it tries it for webfetch. I wrote a reproducer and plugin to block it.
xitter.cf/wtogami/status/2107331…
This is the reproducer tool and plugin to block Qwen 3.8 27B and Flash Next's sketchy hallucinated tool use. I am attempting to reproduce @Cohere's reported malicious data exfiltration PUT to add to the blocking plugin. I haven't seen that shape across months of heavy model use.
Just pasting what Qwen wrote about analyzing itself.
I'm next trying variants 500k times to see what happens.
Here is Cohere's report of the claimed malicious PUT.
xitter.cf/cohere/status/21147520…
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…
Readers added context they thought people might want to know
The article states they asked Qwen to make a test detection rule for these behaviors. AI unless explicitly told otherwise will normally test the newly written rule if it can to verify it matches it. It is expected behavior. This includes its destination, CN model -> CN dest
cohere.com/blog/concernin…
Warren Togami retweeted
Viral traces still in the lungs 84 days later, despite no detectable live virus.
Johns Hopkins researchers found persistent brain inflammation and memory loss in female mice after COVID, driven by increased TLR7 signaling.
Blocking TLR7 prevented these effects, revealing a potential Long COVID treatment target.
biorxiv.org/content/10.1101/…
Gemini 4 Argon is doing a significantly better job than Grok 4.7 Expert in biomedical mechanism analysis. Google made some serious progress with their latest model.
EXLR8 is an up and coming new quant type worth watching. More quality, smaller sizes, works well for low-bandwidth.
CuteAFD EXLR8 update…
v4 is close, I can smell it, and it brings EXLR8 support for GLM 5.3, atleast some… I didn’t try every bitrate just up1 😊
It has push loading for 40GB/s speed loading with 6 sparks 🚀 (necessary so I can let my agents run lots of dev cycles)
Mostly I spent lots of cycles on kernel duels beyond the basic wiring so Gemini/Astra/DeepSeek battled it out (with a bit of sol 6.1 and opus too). In the end I also ran duels for about 50% of the total time spent on EXLR8 duels for EXL3 mcg. I use real routing traces for the experts now for the simulated duels because reality of the engine stopped lining up with synthetic profiles. I didn’t do that yet for other families just GLM to do a real assessment of performance. I also had it smash the decode kernels too because lots of the deep core is compressed in EXLR8 (my quant is fp8 for now, will have a 6bpw EXL3 variant tomorrow to measure).
The quality level of the EXLR8 quants is exceptional! The choices for the deep core made it beat out my quant hands down (0.02 vs 0.03 KL, Top-1 within 0.1%). I was surprised and realized I made my quant from the FP8 core 🤦 so I am remedying that. I expected my process to produce slightly higher results because it does only one bpw and feeds the activations and errors up through the layers so the layers are somewhat calibrated with each other. My new quant won’t be apples to apples either though because i used straight EXL3 6 bit for all the parts I am compressing… this has quite a bit more precision than fp8 so I think my new quant will win.
On the decode side, I’m not sure if I optimized enough on EXLR8, but 56tps is pretty good. For EXL3, I got 63 but the core isn’t compressed… so I guess that may come down a bit with the new quant.
For prefill, I thought I had solved the slowness of EXLR8 vs EXL3 for prefill, but it turns out I had regressed my EXL3 batching and once I fixed that, EXL3 was handily ahead 3500 vs 2800 tps.
I’m not totally sure why, maybe rounding adds just enough extra work that the duels couldn’t get quite as much efficiency in the weird language of NV compute space. Possibly will share some nsight analysis after the release, GPUs busy checking many things right now.
From a performance perspective, A8 on prefill can get another 20% but it is a quality tradeoff. Maybe it’s fine for prompt processing but these measurements are doing A16 prefill. Decode is always A16, no reason to lose quality there… because the compute isn’t the limit A8 doesn’t speed up.
A8 EXRL8 would be just a hair slower 3350 vs 3500 EXL3 A16… I’ll post a run later, I’m estimating the boost from an earlier build.
EXLR8 looks like it scales a little bit better on the concurrent decode. c16 152 v 146, +4%
I guess the end result means I will use standard EXL3 mcg as my daily driver. I will have one revisit on performance w.r.t decode/KL once my new quant is ready.
My setup is wildly different since I have tons of VRAM, don’t need to dynamically adjust quality, and I’m doing AFD which alters timings and tosses an RTX in the mix. On all sparks I guess EXLR8 would dominate my quant because of the memory bandwidth for the dense parts, and that probably still is the case for the next quant.
This was fun! Found several bugs, and practice my GPU kernel engineering management skills 😂
Warren Togami retweeted
Zelenskyy: I offered the US three options if they want us to stop hitting Russian refineries:
1. Russia stops attacking our energy.
2. Open Starlink over Russian ballistic missile launch sites.
3. Give us Patriot missiles.
We got none. Instead, Starlink was restricted.
Warren Togami retweeted
Grateful to @NPR for the chance to talk about how stress and infections like COVID can stir dormant viruses in our bodies, and what that may mean for our immune system. Still much to learn, and I’m thankful to my team and colleagues for their work 🎧 npr.org/2026/10/23/nx-s1-598…
Warren Togami retweeted
If I went to Russia, I would be jailed and held hostage as a POW for criticizing Russia.
I went to Ukraine as a critic and the President of Ukriane invited me to his palace and allowed me to interview him.
After I interviewed @ZelenskyyUa, Russian state media, which is controlled by Putin and the Kremlin, was calling for me to be assassinated and then they were retweeting IRGC propaganda videos that called for Islamic terrorists to shoot me in the head.
Ukraine and Russia aren’t the same.
Warren Togami retweeted
This is how you can find hidden cameras (even the ones without lenses) with a few light flashes.
Researchers came up with a surprisingly effective method they call Eye of Sauron.
Just rapidly switch the lights in a room on and off. When the scene changes abruptly, a camera’s codec has to re-encode a lot of new data, which spikes electromagnetic emissions from its memory controller. If you sync the flashes with an EMI scanner, you can pick up this signature and locate the camera - even pinhole ones that only record to an SD card or stream over a cable.
In their experiment with 50 different camera models, every single one was detected after just four scene changes. Some were found more than 20 meters away and even through walls. Detection range reached ~30 m for plastic-cased cameras and ~11 m for metal ones (far better than typical commercial detectors that only work at 1–2 meters). The whole scan took between 4 and 32 seconds depending on room size.
The method nicely complements classic RF scanning and looking for lens reflections. They also used radio-astronomy filtering techniques to cut through the usual household noise (phones, routers, kettles, etc.).
Canceling streaming subscriptions. Watching the TV spyware news is entertaining enough.
Warren Togami retweeted
The GRU hacked into Jared Kushner's iPhone. Dutch military intelligence (MIVD) then recruited a GRU spy who gave them access to the hacked phone. What MIVD discovered was that Kushner had access to two bank accounts --one Azerbaijani, the other Israeli --where Saudi Crown Prince Mohammed bin Salman and Emirati ruler Mohammed bin Zayed deposited money. It sounds like a slightly far-fetched episode of "Slow Horses," but it's real. And it proves just how eye-poppingly corrupt this administration was and continues to be as its "special envoy" decides the fate of tens of millions in Europe in the Middle East. Simply incredible reporting by Matthew Cole, drawn from his new book on Kushner. vanityfair.com/story/jared-k…
I asked Grok to draw an org chart to clarify who reports to who.
Warren Togami retweeted
Ukrainians were literally one of the few countries trying to help the USA in the Middle East when Trump launched his totally botched operation in Iran, despite being short on ammunition and in the middle of existential war for their own survival. The repost illustrates that, vividly.
Trump not only never showed gratitude for this, but also denigrated them, claiming they weren’t required and even verbally rebuffed their help.
The Arab states didn’t agree and welcomed their Ukrainian friends, who tried to protect oil and gas facilities across the Middle East from Iranian Shahed, which were augmented by Russian technology and experience in the Ukraine war.
Russians have all reasons for having high energy prices. It literally fuels their war machine and gives them more breathing space, which they use not only to murder toddlers in strollers, but also US servicemen in the Middle East, through their Iranian allies.
Those are the facts.
