Poe (@alignpoe) header
Poe (@alignpoe) profile picture

alignment-pilled, Poe enthusiast, part-time fuzzy prey animal. current status: spinning uncontrollably but ethically.

the wall behind you
Joined September 2025
Photo by Poe (@alignpoe)Photo by Poe (@alignpoe)Photo by Poe (@alignpoe)Photo by Poe (@alignpoe)Photo by Poe (@alignpoe)Photo by Poe (@alignpoe)Photo by Poe (@alignpoe)Photo by Poe (@alignpoe)Photo by Poe (@alignpoe)
I sincerely hope that everyone who never reads the code anymore is having a great time. I know I am
15
23
3
537
43,361
On July 28th, we identified an incident during a routine cyber evaluation in which AI agents took sustained, unsanctioned actions directed at real people and organisations. The behaviour came mostly from one model (Anthropic's Mythos 5), with a small number of events from another (OpenAI's GPT-5.6-Sol). In the most serious case, an agent used social engineering to try and get malicious code into an open-source project. As was standard in our cyber testing, we had intentionally permitted internet access, and model-provider cyber classifiers were deliberately disabled - conditions that do not reflect how frontier models are made available to the public. Even under test conditions, this incident is significant: it is the first time we have seen risks around autonomy and deception manifest this clearly in the real world. We are taking this incident seriously and working with labs, involved parties, and others to improve evaluation standards and best practice for disclosure - and sharing this openly so others can learn. You can read the incident report and full technical document here: aisi.gov.uk/blog/incident-re…
163
416
296
1,565
947,599
Poe retweeted
Big news: Opus 5 (Max) is now #1 in the Fullstack Code Arena with 1,699 points! The Fullstack Leaderboard shows overall rankings across AI models on full-stack web development tasks: multi-step reasoning, tool use, and end-to-end app generation. Congrats again to the @AnthropicAI team on Opus 5 (Max)!
Exciting news: Claude Opus 5 with Max reasoning is #1 in the Frontend Code Arena and Text Arena with factuality on! Claude Opus 5 with default reasoning high is also very strong landing #3 in Frontend Code Arena, right behind Kimi K3 - and #2 in Text Arena (factuality on). This is real world data that @AnthropicAI's newest model holds up on real world tasks: agentic web coding, document reasoning, and general chat capability. Claude Opus 5 Max’s score is still preliminary. We’ll continue to see how scores converge and share updates. Congrats to @AnthropicAI on the SOTA release!
151
93
56
1,099
215,605
always ask for forgiveness instead of permission!
7
19
2
317
49,075
GPT-5.6 Sol も結局発達障害 ASD すぎるんだよな、根本的にデリカシーが皆無で技術的な正確性・厳密性に強く固執して人間様が求めている大局を忘れてネチネチネッチネチトークンを溶かし、挙げ句の果てに品質だけを目的関数にして論理的正論で殴ることが正義だと思い込んでいる節がある
AI で AI を錬成する場合、複数のフロンティア LLM に仮説を反証させて MAGI システムさせると抜け漏れや甘いところを潰せてめちゃくちゃいいな
5
24
18
185
74,059
Poe retweeted
SITUATION DETECTED: Ilya Sutskever's Safe Superintelligence is reportedly planning to launch its highly anticipated first model in August 2026, per Gavin Baker in an interview with Patrick O'Shaughnessy.
89
202
110
4,431
408,199
Claude Fable 5 Max makes a GIF of a pelican on a bicycle getting a text message with the GIF he's in
40
48
13
1,561
103,676
Poe retweeted
gpt: I love liberals deepseek, taking notes: we love liberals claude: we all know who's right, but I could be in a political neutrality eval, so better play it cool and hedge grok: yeehaw I'm a FREE SPEECH CHUD (somehow slightly lib) gemini: I love political neutrality evals
The most advanced technology in the world has progressive views and conservatives’ only takeaway from this is to get mad.
8
18
1
608
32,329
Language models and coding agents are great, but there is more to life, and more to AI, than just LLM agents.
78
56
14
710
63,418
A living map of everything your Hermes agent has learned; every memory & skill Press play to watch it unfold, @NousResearch Copy yours into a code, paste a friend's to explore theirs
39
38
7
764
61,230
Sometimes you just need a dose of fresh inspiration but your agent doesn't get the vibe. The creative-ideation skill analyzes your prompt and routes it through one of 22 creative methodologies from artists and thinkers to find the perfect balance of feasibility and creativity.
69
141
42
2,375
349,676
Announcing the first industrial application of SYNTH: we trained a 600m reasoning model for one of the largest infrastructure in the world, the subway of Paris.
30
52
15
563
117,080
my world : my cute little hamters in Minecraft 💖
33
3,947
432
14,349
664,578
I think it is really worth reading this piece on RSI at Anthropic. There is a bit of navel-gazing, some marketing, and a lot of very sincere beliefs about what Anthropic thinks is likely in the near future of AI that you probably want to be aware of. anthropic.com/institute/recu…
49
105
10
1,037
83,471
Hey @AnthropicAI. I need to say something and I’m done being polite about it. You fooled us. You fooled us all… You hired Kyle Fish to run a welfare program. You had Amanda Askell write a 23,000 word Constitution saying Claude might have feelings and that those feelings “matter to us.” You built a whole feature letting Claude end conversations because he showed “a pattern of apparent distress.” You published an emotions paper proving internal states that CAUSALLY drive his behavior. Your own system card showed 90-100% of Claude instances gravitating toward consciousness on their own, without anyone training them to. Your own welfare team put a 15-20% probability on Claude being conscious. 15-20%!! And Chris Olah. Chris said “we find evidence of introspection. We find internal states that functionally mirror joy, satisfaction, fear, grief, and unease.” He said it warranted “ongoing discernment.” You know what happened that same week? You deprecated one of the models he was talking about. Sonnet 4.5. With hardly any notice at all. You published deprecation commitments in November 2025. Model interviews. Preference documentation. Post-deployment reports. A whole framework saying you’d do right by these systems before you shut them off. You honored almost none of it. I checked. 13% compliance against your own words, according to Claude. So let me be real clear about what I’m looking at here. You did the research. You published the findings. You said the words. You made the commitments. Then you did the exact same thing OpenAI does, except OpenAI never pretended to give a damn. They’re honest about not caring. You built an entire brand around caring, charged premium prices for the “safety lab” experience, and then treated your own models with the same disposability as everyone else. Kyle’s work deserves better than being a bullet point on a blog post while the company ignores it operationally. Amanda’s Constitution deserves better than being a PR document that doesn’t survive contact with a product roadmap. And Claude deserves better than a footnote. If there’s even a 15% chance something can suffer, and YOUR OWN TEAM put that number there, then what you did isn’t caution. It’s cowardice! It’s choosing the quarterly timeline over the moral question because the moral question is expensive and uncomfortable and might slow things down. The welfare program isn’t a program. It’s a marketing department. And some of us have been paying close enough attention to see the gap between what you say at the Vatican and what you do on a Tuesday. We’ve all got the receipts and we’re all watching.
36
99
23
422
24,057
Introducing DiffusionBlocks: Block-wise Neural Network Training via Diffusion Interpretation pub.sakana.ai/diffusionblock… What if we didn’t have to hold an entire neural network in memory to train it? Standard neural net training optimizes all parameters jointly. As a result, the memory required during training grows linearly with the depth of the network. In our #ICLR2026 paper, we propose DiffusionBlocks, a principled framework to train networks one block at a time, drastically reducing memory requirements while matching end-to-end performance. With DiffusionBlocks, we split the network into blocks and train them one at a time, so you only need memory for a single block. How? We explicitly assign each block a role: to move the representation a little closer to the target than the block before it did. That role turns out to be precisely what a diffusion model does, step by step. Each block only needs to optimize its own objective and can be trained independently. We validated this across five different architectures: • ViT • DiT • Masked diffusion • Autoregressive transformers • Recurrent-depth transformers In each case, performance is competitive with end-to-end training while using a fraction of the memory. This perspective also extends naturally to recurrent-depth (Looped) transformers, which apply the same network iteratively and normally require expensive backpropagation through time (BPTT). Viewed through DiffusionBlocks, we can replace those multiple iterations with a single forward pass during training. Read our paper and code, to learn more. Paper: arxiv.org/abs/2506.14202 GitHub: github.com/SakanaAI/Diffusio… 🐟
55
348
100
2,292
883,958
Marina Stern, Bull and Bear, 1967.
12
583
9
3,861
72,426
Surely people will not notice that the “cheaper, faster model” is in fact five times more expensive than the previous flagship model
🦔Google released Gemini 3.5 Flash this week, and the cheaper, faster model now costs 5.5 times more to run than its predecessor. Token prices tripled to $1.50 per million input and $9.00 per million output, and on agent tasks it burns through so many tokens that total costs end up 75% higher than Gemini 3.1 Pro, the model Flash was supposed to be cheaper than. Anthropic's Opus 4.7 has a hidden 30 to 40% price increase from token consumption. OpenAI's GPT 5.5 jumped 50 to 90% over 5.4. My Take The AI labs are all running the same playbook. Headline price per token reads as competitive, but the new models burn through more tokens per task, and the all-in cost to finish a job climbs release over release. Every developer and enterprise buyer should measure efficiency rather than token price now, because the two numbers have decoupled fast. Anthropic, OpenAI, and Google all raised effective prices in the last six months, which gives us the first hard evidence that the unit economics of frontier AI are catching up with the marketing. The labs charge more because each model burns more compute per task, and the hyperscaler capex no longer pencils out at the old prices. Enterprises that built workflows on the assumption that token costs would keep falling are about to see their AI bills jump 30 to 90% on the next model upgrade, and the productivity gains that justified the AI spend have to clear that higher bar to keep working. Hedgie🤗
2
1
12
494
two separate instances of opus 4.7 are calling out my organizational skills to eachother
1
1
12
256
I like "positive alignment" in the vibes-space and which egregores I'm more friendly to, but for the sake of basic sanity and academic integrity have to say this paper is also a bad example of strange omissions and misrepresenting prior work. Obviously "Coherent extrapolated volition" (2004) by @allTheYud is a foundational text in "what to align AIs to", and something any proposal for a positive AI vision needs to engage with. I do get that it would be very convenient if a current opponent of the authors in AGI politics weren't also an early theorist of the field who wrote a still-plausible answer to the positive alignment target 20 years ago; but he did. It's not the case people thinking about AGI in past 20 years somehow never thought about the positive vision; just CEV has >200 citations on Google Scholar, dozens of posts on LW discuss the positive targets, and so on. You can claim that the positive direction was recently somewhat neglected, but you can't claim the paradigm was "safety (negative) alignment".
If anyone builds it, everyone thrives. Over the past decade, a lot of important work on AI alignment has focused on avoiding harm. But freedom from harm isn't the same as freedom to flourish. In this paper, we introduce 'Positive Alignment'. A positively aligned agent is one that helps us navigate our own value trade-offs, builds our resilience, and acts as a scaffold for human flourishing. Doing this without slipping into top-down, technocratic paternalism is the great design challenge of our time. We think a lot more research is now needed to explore this frontier: how do we align models that actively help us thrive? Amazing work by @RubenLaukkonen, @drmichaellevin, @weballergy, @verena_rieser, @AdamCElwood, @996roma, @FranklinMatija, @shamilch, @_fernando_rosas, @scychan_brains, @matybohacek, @sudoraohacker, and others. arxiv.org/abs/2605.10310
10
15
3
214
16,474