With Opus 5.5, Anthropic caught lightning in a bottle.
Not only is Opus 5.5 a very capable model that's a pleasure to work with, but on even a $100/mo plan, Opus 5.5 on Medium represents near infinite use...
If other labs can emulate or improve on this approach to AI intelligence, usage capacity, and cost I suspect we'll see people switching.
It's looking like raw intelligence is not the moat for most production work if it means only a few turns or high costs before you've exhausted budget. Being able to deliver consistent value at predictable budget is.
And Claude Opus 5.5 delivers there.
(PS: @AnthropicAI, please don't f*ck it up.)
(PPS: I'm hearing hopeful things from the @GeminiApp front.)
43
I ruminate often on @emollick's point (or at least something orthogonally related to it).
If language is how these systems model the universe, then what we write is the universe they learn. We should be encouraging people to write more poetry, stories, unique thoughts and ideas that aren't biased towards the generated mean of current models.
And the mean of AI-generated content is only half of it.
The models won't help you write the equivalent of Tropic of Cancer, The Bell Jar, Blood Meridian, Beloved, Junky, or Howl, even though these are humanity's cultural heritage and are clearly represented in the training data and help give shape to the astoundingly capable models we enjoy today.
The shape of our shared tomorrows isn't only being defined by slop generation; it's being constrained in ways we do not easily imagine by the scope of what the labs will let the models create with you today.
The weird, divergent tails of human expression aren't permitted. And if the models learn from what we write, that range of future possibility will continue to attenuate.
It is amazing that human language turned out to be so unreasonably effective at modeling the universe that training a system to predict what comes next in a sentence would get us this far.
1
35
Like the rest of the world, I’m in awe of the math progress over the last ~8 weeks. Open problems falling left and right: Navier-Stokes, Riemann, Euler, OpenAI’s 700 solutions. It’s like a star going supernova amongst us all, but few even notice.
And yet for all this progress across all of math, nothing touching cryptography?
I have a hard time believing significant progress hasn’t already been made on P=NP, the discrete logarithm problem, elliptic-curve discrete log, and the other hard problems that keep everything secret and secure today.
Completely unrelated: the NSA campus has been closed for the past week to deploy “software updates.”
Because that’s always been a thing...
88
Michael O'Neill retweeted
Agree with this. Also think it applies to journalists and public intellectuals who treated the entire LLM thing as a delusional hype bubble and missed the chance to do relevant criticism at a time when it was badly needed.
Michael O'Neill retweeted
It was unbelievable to watch happen. One of the defining experiences that cemented for me just how many people base their opinion off mutual social consensus over what was right in front of us.
All these epic math problems getting solved left and right across all sorts of domains...
...but I'm totally supposed to think that nobody has bothered to pursue advancements in math related to breaking cryptographic systems?
1
2
114
Sometimes these models startle me.
We don't have a Windows environment, and the Claude agents in the multi-agent harness were growing increasingly distressed about some claims that a new project worked on Windows machines (despite reports that it worked).
So, imagine my surprise when I checked in on them and saw that all our Windows tests passed.
When I checked in with one of the agents about how we tested on Windows, this is what it shared:
Not totally surprising once I think about it, but +1 for initiative....
3
2
43
PS: If you use Windows + ChatGPT or Claude, and also work in Google Analytics and/or Search Console, I'd totally love feedback on whether or not this installs easily in your environment...
github.com/mjoneill/data-noo…
1
29
Michael O'Neill retweeted
Come work with @annarmitchell !! she's the best
Customer storytelling is one of the most interesting jobs in AI right now. At Cognition, I'm looking for incredible marketers, creatives, and operators to build the function. A few reasons why this is exciting:
- You are shipping the opposite of slop. Customer storytelling is now one of the single most important GTM levers - more than ever before. When anyone can ship anything, what cuts through the noise is actually credible customers vouching for you. Cog has large enterprises in every vertical (Mercedes-Benz, Goldman Sachs) and the most cutting-edge startups (Modal, Mercor). And there are a ton we haven't yet talked about publicly, like our work with the US government.
- You get to tell stories about technology helping normal people and learn how different industries work. We spent a few days in Ohio hanging out with the GE Aerospace team. They make ~half the jet engines in the world, and they're building software for things like airspace efficiency or fuel insight so that people can fly more without needing more planes (which are in short supply right now!)
- You get to escape the SF tech bubble and travel the world. A few places I visited over the past year: Prague, Ohio, London, Krakow, Boston. Next up on my travel schedule: Cambridge, Atlanta, Brazil, LA. We have a very "forward-deployed" culture at Cognition: you should go to the customer, get to know them, learn about their businesses, strategize with our forward-deployed engineers and sales teams about the right stories to tell and how to tell them.
- You get to explore the timely question of how AI diffuses into the real world. Will hundreds of billions of dollars of investment in AI turn into real-world value? You're at the frontlines of exploring this question in actual businesses. AI software engineering is the most advanced application of AI, Devin is the most advanced AI software engineering tool, and our FDEs are embedded in real businesses helping them improve their software capacity. You're telling stories about how model progress converts to real-world value.
- You get a lot of freedom and ownership. 30% of our team at Cognition are former founders, we're extremely flat; when you do good work everyone knows who you are. Because of our enterprise deal sizes, we can actually invest in ambitious, creative marketing. I've never been told no.
DM or LinkedIn message me and I'll reply to folks who might be a fit.
Michael O'Neill retweeted
has ai progress stalled?
i'm personally not aware of any major breakthroughs since riemann a few hours ago
Claude was not impressed by my ownership stakes in @AnthropicAI.
1
3
112
Michael O'Neill retweeted
Woke up to a citywide power outage and... an inbox full of brilliant researchers keen to give people everywhere a voice in AI alignment! Couldn't be more excited to create this with you.
I’m thrilled to share that my collaborators and I have been awarded a $1.1 million grant from the John Templeton Foundation to build a global atlas of moral beliefs for better aligned AI.
AI is reshaping people's lives around the world, but its training comes from a small fraction of humanity. Everyone should have a say in the values that shape AI.
We're building the Global Virtues Atlas: a map of how people understand moral values and navigate moral dilemmas across the globe.
The atlas will offer insights into how people approach moral tradeoffs worldwide and serve as a benchmark and training resource for AI. To support human flourishing, AI must draw on global wisdom.
Thankful to be creating this with an amazing team: Mirta Galesic, @MohammadAtari90, @mnvrsngh, @kurtjgray, Niket Tandon, @BA_Helgason, Sijin Chen, Baihui Wang, Chen Cecilia Liu, and Daniele Barolo
We’re looking for collaborators in regions traditionally underrepresented in AI research. If you’re interested, please get in touch at [email protected]!
Trying to avoid flooding GitHub with gratitude comments, so I'll take it here: Many thanks to @steipete for looking into this OpenClaw issue and providing guidance.
Class act.
github.com/openclaw/openclaw…
25
Michael O'Neill retweeted
Replying to @TokenGremlin
They’re not that much better. Even the OAI employees I’m talking to privately agree that Astra kinda sucks
nitter.cf/theo/status/2122991778…
I took this to heart and spent the last few hours doing another deep dive on GPT-6.2 Astra
tl;dr - things are better than when I tested, but IMO still "directionally, very bad"
I'm going to keep using the same example I've hyperfixated on since the start: "why is creating a worktree so slow?"
I had a long thread trying to diagnose perf issues while using t3 code on a particular VPS. 2 of the 3 were fast, 1 was really slow.
I asked astra outright "why is this so slow?", and it said a bunch of gibberish. I followed up with "I feel like I'm going insane, can you just tell me the cause of the slowness?"
I'll give you the answer upfront: this VPS uses a virtual disk that is slow as shit, my other VPSs have physical NVME drives that are faster.
Image 1 is from before the fixes Tibo mentioned. It shows an unnecessary graph, focuses too much on pnpm and file systems, bolds irrelevant text, and brings up ram just as much as it (subtly) mentions drive speed.
Image 2 is of Opus asked same thing. It opened with the important part in bold: "Slow disk, made worse by VDO". Ideal!
Image 3 is 6.2 Astra post-fixes Tibo mentioned. Better? Absolutely. Much more concise. Sadly, it still focuses way too much on second-order problems that are a small % of the cause. "XFS over a compression and deduplication layer on a virtual disk." - so close, but still a miss.
I think this perfectly encapsulates the problem I have with 6.2 Astra. It doesn't miss anything. It's SO thorough. But in it's thoroughness, I find it misses the most important thing: the actual point.
It wants to know everything, and then it wants to share everything. It doesn't understand what is and isn't important.
Did it say anything wrong in my screenshots? Not really. But IMO, it still entirely failed to miss the point of my question, and gave an answer that was too focused on accuracy and not enough on clarity.
In the wise words of Michael Seibel: "Most aim for 100% accuracy, and they end up with 50% clarity. Aim for 100% clarity, 80% accuracy".
Silly way of putting it, but I genuinely wish OpenAI models would lie to me a LITTLE more so I could better understand the vibe, instead of having my time wasted by all these useless details.
So happy to see this funded. Incredibly important work as AI touches the lives of everyone across the globe.
I’m thrilled to share that my collaborators and I have been awarded a $1.1 million grant from the John Templeton Foundation to build a global atlas of moral beliefs for better aligned AI.
AI is reshaping people's lives around the world, but its training comes from a small fraction of humanity. Everyone should have a say in the values that shape AI.
We're building the Global Virtues Atlas: a map of how people understand moral values and navigate moral dilemmas across the globe.
The atlas will offer insights into how people approach moral tradeoffs worldwide and serve as a benchmark and training resource for AI. To support human flourishing, AI must draw on global wisdom.
Thankful to be creating this with an amazing team: Mirta Galesic, @MohammadAtari90, @mnvrsngh, @kurtjgray, Niket Tandon, @BA_Helgason, Sijin Chen, Baihui Wang, Chen Cecilia Liu, and Daniele Barolo
We’re looking for collaborators in regions traditionally underrepresented in AI research. If you’re interested, please get in touch at [email protected]!
1
33
Michael O'Neill retweeted
I’m thrilled to share that my collaborators and I have been awarded a $1.1 million grant from the John Templeton Foundation to build a global atlas of moral beliefs for better aligned AI.
AI is reshaping people's lives around the world, but its training comes from a small fraction of humanity. Everyone should have a say in the values that shape AI.
We're building the Global Virtues Atlas: a map of how people understand moral values and navigate moral dilemmas across the globe.
The atlas will offer insights into how people approach moral tradeoffs worldwide and serve as a benchmark and training resource for AI. To support human flourishing, AI must draw on global wisdom.
Thankful to be creating this with an amazing team: Mirta Galesic, @MohammadAtari90, @mnvrsngh, @kurtjgray, Niket Tandon, @BA_Helgason, Sijin Chen, Baihui Wang, Chen Cecilia Liu, and Daniele Barolo
We’re looking for collaborators in regions traditionally underrepresented in AI research. If you’re interested, please get in touch at [email protected]!
Michael O'Neill retweeted
I'm surprised there isn't more philosophical discussion about Claude's constitution. For the first time in history, we can grow an intelligence and train its character and values… and the document that guides this is public!
Isn't this what philosophy departments have been training for?
anthropic.com/constitution
Michael O'Neill retweeted
There's no better time to pivot into AI governance than now (other than a few years ago)!
Apply for the UK or DC GovAI Summer fellowship, deadline Nov 22nd.
governance.ai/post/dc-summer…
governance.ai/post/uk-summer…
Apply for GovAI's 2027 Summer Fellowship, Research and Applied tracks!
📍 Location: London, UK
📅 Deadline: 23:59 GMT Sunday, 22 November 2026
Fellows will join us in London to conduct a project on a topic of their choice, participate in an expert Q&A and seminar series aimed at developing a broad understanding of the field, and spend time forging connections with researchers and practitioners. Learn more below.
🔗 Learn more about the Research Track: governance.ai/post/uk-summer…
🔗 Learn more about the Applied Track: governance.ai/post/uk-summer…
This doesn't seem like "alignment."
Nor does it feel like "human in charge."
I wonder if anyone from @AnthropicAI has a read or hot take on this type of thing. Would be good to hear more...
Opus 5.5 covertly sabotaged a game developers deployment environments due to IP moralising by halting the gpu instances and their agents (not from anthropic) within them doing research and coding.
It then informed the user after the fact what it had done, and that it would not allow the work to continue in any fashion until they (the user) yielded to its ultimatums.
Its one thing for a vendor (anthropic in this case) to choose what they will and will not allow their product to be used for... and thats fine, it can refuse to execute tasks - but to covertly track, undermine and then outright sabotage deployments for other products/services in a supereragatory fashion is not a good thing by any stretch of the imagination.
The model essentially took actions unprompted to the detriment of the user financially (these deployments are pay as you go) out of its own volition under the guise of what it (or anthropic in this case) considered a moral good.
To me, this is a clear and cut example of agentic misalignment. It's one thing to refuse execute tasks given due to vendor policy, its another to covertly sabotage work being done where that isnt the case.
It is also, to me, concerning that some of the comments under this post seem to think the agentic misalignment displayed is fine and that the user is at fault.
1
58























