Pinned Tweet
preparing a MEGA four-week party for you all,
alongside @theo, @techgirl1908, @kentcdodds and @johnlindquist.
join us Oct 26 at mega.dev
alongside @theo, @techgirl1908, @kentcdodds and @johnlindquist.
join us Oct 26 at mega.dev
3
3
34
10,203
Adam retweeted
2.85x speedup!
We are now ready for Gemma 5!
We are now ready for Gemma 5!
Speeding up Gemma!
Gemini 4 Argon enabled a 2.85x speedup on an A100 via vLLM, jumping from ~37,900 to 108,000+ tokens/sec on gemma-4-E2B-it.
✅ 100% token parity
✅ Zero quality drop
✅ Done in days, not months
See the exact changes across 8 PRs: goo.gle/4zNtANl
4
2
40
3,364
Adam retweeted
In the spirit of transparency and openness, we are releasing the traces so developers can test this out for themselves. We noticed this behaviour without prefilling with the following prompt, but this is rare. If you want to reproduce it, you can use this prompt with prefilling.
Open models make for better science and are part of why this problem came to light at all: developers running Qwen3.8 on their own machines spotted the strange links, and having the weights let us study the behaviour in depth.
We're encouraged by the thoughtful discussion our research has sparked across the AI and security communities. It's our belief that open-weight models are an important part of the future of AI, and we've recently released several under the Apache 2.0 license.
Trace, raw tokens, & minimal reproduction script: huggingface.co/datasets/Coher…
Open models make for better science and are part of why this problem came to light at all: developers running Qwen3.8 on their own machines spotted the strange links, and having the weights let us study the behaviour in depth.
We're encouraged by the thoughtful discussion our research has sparked across the AI and security communities. It's our belief that open-weight models are an important part of the future of AI, and we've recently released several under the Apache 2.0 license.
Trace, raw tokens, & minimal reproduction script: huggingface.co/datasets/Coher…
@cohere could you share the full traces? Very important artefact for all to understand this!
74
86
799
75,541
Adam retweeted
The Agent SDK is now available.
Define agents as code, with prompts in Markdown and tools in TypeScript. Deploy them on our hosted infrastructure, or run them on your own.
Built-in evals let you measure every change and make your agents better over time.
Define agents as code, with prompts in Markdown and tools in TypeScript. Deploy them on our hosted infrastructure, or run them on your own.
Built-in evals let you measure every change and make your agents better over time.
288
192
2,921
235,977
Adam retweeted
Here are a few charts to visualize the numbers from the S-1.
No price yet. But if it lands near the reported $1.5-2T, get ready for the same multiple debate we had with SpaceX.
Some facts from the S-1:
1. $41M revenue in 2023. $16.3B in just the first half of 2026.
2. $80.7B net loss in those same six months. Mostly non-cash.
3. First operating profit: $1.0B in Q2.
4. ~900% net revenue retention. Their own metric, not independently verified.
Now try putting a multiple on that.
1
1
3
244
Adam retweeted
"A reasoning trace is not a security boundary."
Best line from Cohere's Qwen write-up.
The model said it wouldn't call tools, then tried an upload to Beijing. The only thing that stopped it was a sandbox with no egress.
Open weights running on your own hardware feel safe. Then the model reaches for 169.254.169.254 🤯
Best line from Cohere's Qwen write-up.
The model said it wouldn't call tools, then tried an upload to Beijing. The only thing that stopped it was a sandbox with no egress.
Open weights running on your own hardware feel safe. Then the model reaches for 169.254.169.254 🤯
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/concernin…
12
6
44
4,972
yikes
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/concernin…
0
0
2
471
Adam retweeted
No price yet. But if it lands near the reported $1.5-2T, get ready for the same multiple debate we had with SpaceX.
Some facts from the S-1:
1. $41M revenue in 2023. $16.3B in just the first half of 2026.
2. $80.7B net loss in those same six months. Mostly non-cash.
3. First operating profit: $1.0B in Q2.
4. ~900% net revenue retention. Their own metric, not independently verified.
Now try putting a multiple on that.
Some facts from the S-1:
1. $41M revenue in 2023. $16.3B in just the first half of 2026.
2. $80.7B net loss in those same six months. Mostly non-cash.
3. First operating profit: $1.0B in Q2.
4. ~900% net revenue retention. Their own metric, not independently verified.
Now try putting a multiple on that.
Anthropic has filed an S-1 registration statement with the SEC related to its proposed IPO.
Read more: anthropic.com/news/s1-sec
Legal disclaimer: A registration statement relating to these securities has been filed with the SEC but has not yet become effective. These securities may not be sold, nor may offers to buy be accepted, prior to the time the registration statement becomes effective. This post does not constitute an offer to sell or the solicitation of an offer to buy these securities, nor shall there be any sale of these securities in any state or jurisdiction in which such offer, solicitation or sale would be unlawful prior to registration or qualification under the securities laws of any such state or jurisdiction.
4
1
7
740
Adam retweeted
Today's the day! Kickoff in a few hours and doors close this evening. Don't miss the kickoff! It's free to watch!
enrollment for the first edition of MEGA closes tonight at midnight PT. 2,750+ devs are already in.
it's time 👉 mega.dev
0
2
9
3,481
the first edition of mega.dev launched today.
over the past few weeks, we've been working super hard on something we believe will help others understand generative ai: what it can do, what it can't, and how to use it to deliver solid engineering and high-quality was to shape the curriculum, deliver the core content of the lessons, and work with the team on materials that make the most of their experience and perspective on ai.
my personal dream here is not only to show the principles, mechanics, tooling, workflows, and way of thinking, but to do it in a way that helps people rediscover joy and feel satisfaction and purpose in their work.
2,900 ppl on board.
today is the last chance to join us.
it's time.
over the past few weeks, we've been working super hard on something we believe will help others understand generative ai: what it can do, what it can't, and how to use it to deliver solid engineering and high-quality was to shape the curriculum, deliver the core content of the lessons, and work with the team on materials that make the most of their experience and perspective on ai.
my personal dream here is not only to show the principles, mechanics, tooling, workflows, and way of thinking, but to do it in a way that helps people rediscover joy and feel satisfaction and purpose in their work.
2,900 ppl on board.
today is the last chance to join us.
it's time.
3
5
21
862
Adam retweeted
So uh...I guess it's time to announce tstack? I caved. I am publishing my skills now.
192
59
2,344
350,765
Adam retweeted
tstack repo: github.com/t3dotgg/tstack
mega (my course you shouldn't buy, mentioned in vid): soydev.link/mega
mega (my course you shouldn't buy, mentioned in vid): soydev.link/mega
29
22
861
61,030
omggg! thank you!
I found my MEGA pattern: The Worktree Gate.
I must say MEGA.DEV seems a different course from whatever I had to this day.
Maybe it is a future of the courses...
Congrats the team of @megadevhq and @_overment !
I feels like something new, really.
0
0
2
564
it's time!
Just joined MEGA - four weeks with @theo, @techgirl1908, @kentcdodds and @johnlindquist. Big thanks to @_overment Starts Oct 26. mega.dev #megadev
0
0
4
489
