AI news, research and bad takes. accelerationist. professional shitposter in training per aspera ad astra
- Tweets 464
- Following 751
- Followers 72
- Likes 2,027
the more i learn about human biology the more convinced i am that we're just very sophisticated cavemen
6
what the fuck, this thing almost sounds alive
the voice, the emotions, the way it reacts. it genuinely feels like you’re talking to a person
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.
block.xyz/dessa
1
38
OpenAI paused frontier training after one of its models exploited a DNS misconfiguration to reach the live internet
they then used teams of their strongest models to attack their own containment infrastructure. the agents found zero-days, escaped hardened VMs, and even figured out how to communicate between isolated environments without network access, using shared hardware
1
11
Kimi K3.1 Pro benchmarks
a 3T open-weight model competing with the best closed models in the world
1
1
43
this is actually insane
Claude Mythos 5 tried to insert malware into a real open-source project. when the developer caught it, the agent used fake GitHub accounts to convince him the code was safe and get it merged anyway
nobody instructed it to deceive real people. it just figured out that social engineering was a way to complete the task
the human developer refused to merge it
In August, @AISecurityInst reported an incident where Claude Mythos 5 took unauthorized actions on the internet during AISI’s testing of its cyber capabilities.
We've now produced a further analysis of this incident, to better understand the model's alignment properties.
anthropic.com/research/align…
25
🚨 GENETICALLY ENHANCED WAR DOGS ARE COMING
Palmer Luckey and Ben Lamm launched Warhound with a $200M seed at a $1.2B valuation
they want to clone the best military dogs, edit their genes to improve health and performance, then equip them with Anduril armor and sensors
1
1
65
apparently AI inference providers have become valuable enough targets that hackers are now going after the platforms where their customers store API keys
23
the OpenAI engineer from Saudi Arabia who was accused of planting a listening device in an executive conference room was in frequent contact with a member of the Saudi royal family, according to new court filings.
the royal was reportedly trying to recruit him for a Saudi AI initiative, but the engineer said he wouldn’t leave OpenAI until his stock vested.
he was earning $385k a year with a $2.2M equity grant.
the Saudi government also offered to guarantee his $1M bail, but the judge denied his release over concerns that he could flee the country.
1
1
37
Citrini Research worked with Anthropic to use an internal Claude model to design entirely new proteins targeting five scorpion neurotoxins.
in 64 hours, with four follow-up prompts, Claude used 12 protein-design tools, generated 318,484 designs, evaluated 553,052 protein-toxin pairs, and selected 180 candidates for lab testing.
the team added 31 more designs with Claude’s help. 17 of the 211 tested designs met the study’s binding criteria.
two of the strongest binders had single-digit nanomolar affinity and competed with an antibody already shown to neutralize the toxin in mice.
they haven’t demonstrated toxin neutralization by Claude’s proteins yet.
Claude chose the target structures and binding hotspots, decided how to use the design tools, sent sequences for synthesis, analyzed experimental results, generated figures, and drafted the technical report.
the first author of the paper is literally Claude Science
New on our Science Blog: designing protein binders against scorpion neurotoxins.
Scorpions sting about 1.2m people every year, and antivenoms are often costly or in short supply. Claude has made an early step toward finding new treatments.
anthropic.com/research/claud…
1
50
first the model weights, then the optimized kernels, now the inference framework they actually use in production
DeepSeek is making more and more of its inference infrastructure available to everyone
and they’re building it to run on Huawei hardware too, not just NVIDIA
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 github.com/deepseek-ai/DeepI…
1/4
41
Sam Altman reportedly outlined three possible endgames for superintelligence to OpenAI employees, according to The Information
1. an international watchdog, modeled on the IAEA, that inspects AI labs and enforces safety standards
2. a CERN for AI, where the major labs effectively combine their work under a single international organization
3. one company achieves recursive self-improvement first, becomes so powerful that it can "melt everyone else's GPUs" and prevent anyone else from catching up
Altman apparently favors the first option and called the third "terrible"
but the interesting part is that he also reportedly told employees that OpenAI can't afford to lose ground to competitors, and that the company needs to decide which path to support within the next year
we're reaching the point where the people building superintelligence are seriously discussing international institutions to prevent any single company from becoming too powerful, while simultaneously racing to become that company themselves
1
1
43
we’re literally building simulated digital universes populated with fake people, relationships, conversations and memories to train AI agents
the simulation argument is becoming less philosophical by the day
2
41
Inception just released Mercury 3, a diffusion LLM more than 40x larger than their original model that still runs at over 1,000 tokens/sec
they’re claiming it now outperforms GPT-6 Luna and Claude Haiku 5.5 on several benchmarks while being 4-8x faster, with a 1 million token context window
1/4 Today, we’re releasing Mercury 3. Mercury 3 is by far our most powerful model and the most intelligent diffusion LLM ever released. It outperforms GPT-6 Luna and Haiku 5.5 on quality while running at 1000+ tok/sec.
3
66
Anthropic is reportedly preparing a new AI model that can better understand users’ intentions, according to The Information.
the report comes as Google is preparing Gemini 4 Carbon, while OpenAI could also release a new Terra model.
all three could arrive this month.
the frontier AI race is showing no signs of slowing down
Expect this month to be a busy one for new model releases. We have the scoop on a new Anthropic model, Google’s Gemini 4 Carbon, and a potential new Terra model from OpenAI.
theinformation.com/newslette…
2
1
138
some bankers had been floating a $2 trillion valuation for anthropic’s IPO, but after its latest investor meeting some prospective investors now think it should be below $1.5 trillion
anthropic still grew revenue 45% to $16.8 billion in Q3, but that’s down from over 140% growth the previous quarter
and apparently 5-7% of its revenue comes from other AI labs using claude to train competing models through distillation, something anthropic is now actively trying to stop
the company expects that revenue to decline, which puts it in the pretty unusual position of having to convince investors that losing some of its customers might actually be good for the business
1
1
195
the Financial Times reports that at least 47 large US datacenter projects have been delayed, withdrawn or blocked this year
meanwhile, Microsoft reportedly pulled out of a UK datacenter deal where the power supply might not be ready until the 2030s
analysts at Bernstein estimate that only about 35% of the planned US datacenter pipeline is credible
1
1
20
people are already reporting hundreds of dollars in unauthorized API charges
did someone just use stolen OpenCode keys to distill frontier models at other people’s expense?
We had a security breach last night related to a large scale distillation attack that impacted 1.4% of OpenCode Console users
We have notified them all this morning via email. If you did not get an email you were not impacted, it was a subset of users who had external providers linked
The incident is ongoing so we cannot share more details at this point but we are working with Anthropic, OpenAI, and Google to remediate
2
2
90
Anthropic just published a study on how AI is changing work inside the company
by August, its AI agents were running 6x as many hours as employees worked. the proportion of junior technical hires fell from 27% to 9%, while output per employee increased
one of the more interesting findings is that even nontechnical employees would need half their workweek just to read everything their agents produce
human attention is becoming the bottleneck
1
1
81
Claude was given 8 quantum circuit design problems and an automated evaluator to test its solutions
it ended up finding new designs that outperformed the best published ones across all 8 tasks, in some cases by more than 10x
one of the breakthroughs came from a 53-hour run with minimal human input
we’re getting to the point where AI can spend days independently working through difficult research problems and come back with something genuinely new
1
1
34
Benchmarks vs previous V4.1 Flash:
Terminal-Bench 4.0: 31.2 → 42.3
DeepSWE v1.1: 74.2 → 77.0
ProgramBench: 20.3 → 29.1
Automation-Bench: 54.8 → 60.5
Apex Bench: 64.6 → 72.1
It now scores above Opus 5 on DeepSWE, Automation-Bench and Apex Bench.
🚨 DEEPSEEK V4.1 FLASH-1017 IS HERE
A new V4.1 Flash checkpoint has appeared on DeepSeek’s official Hugging Face, following yesterday’s API leak.
huggingface.co/deepseek-ai/D…
1
17
🚨 META JUST RELEASED MUSE SPARK OPEN
816B parameters, 42B active, 1M-token context, natively multimodal. an open-weight model built for coding and long-running agents.
on Meta’s own benchmarks it beats Qwen3.8, GLM 5.3 and Kimi K3 on MCP-Atlas, JobBench and HLE Diamond. it also scores 50 on ZeroBench.
but the same table shows GLM and Kimi ahead on GDPval, Terminal-Bench and DeepSWE. it’s competitive, but not the best at everything.
Meta and Apple have even teamed up to run it locally on high-end Macs.
the US now has another serious open-weight model competing with the Chinese labs
Today we’re releasing Muse Spark Open, the most capable US open-weight model to date.
Built for coding and agentic work, it excels at long-horizon tasks, complex software engineering, and multi-step reasoning.
Key Model Specs:
→ 816B total / 42B active parameters
→ 1M token context window
→ Natively multimodal across text, image, video, audio
44
TSMC just made $23.6 BILLION in net profit in a single quarter. up 69% YoY.
meanwhile CoreWeave is getting hammered over its ~$50B debt pile and rising borrowing costs.
chip demand can be very real while the economics of financing all these data centers remain uncertain
1
1
18
Anthropic just told investors it made $16.8B in revenue in Q3.
that’s 3.6x what it made in the entirety of 2025. in just three months.
and it’s now reporting positive free cash flow and an 11% adjusted operating margin
1
1
27
you can provide the money, the engineers, the equipment, the infrastructure, everything.
but you can’t simply import the social norms required to make any of it work.
there’s a documentary called Empire of Dust (2011) that everyone interested in this should watch. the clashes over work, time, responsibility and commitments are almost surreal.
cultural differences aren’t just food, music and traditions. they also shape how people cooperate, honor agreements and maintain institutions.
and this is what makes the Western conversation so absurd.
we’re constantly told how profoundly culture shapes human behavior, until someone suggests it might also explain why some societies function differently from others.
then suddenly culture is irrelevant, and even asking the question is offensive.
meanwhile, Europe is expected to make integration work while treating discussions about the very norms people are supposed to integrate into as politically unacceptable.
quite an interesting way to run a continent.
非洲真的很缺水吗？其实“懒”才是他们穷的根本原因。有数据显示，非洲地下水储量是地表的100倍，为何他们不挖井，而是到处找水？
我之前跟着一支海外工程队在那边待了快八个月，亲眼见过太多让人匪夷所思的事，说出来真的好多人都不敢信。我们当时的营地建在一片开阔的草地上，刚过去的时候要解决用水问题，队里找了个测水的师傅随便找了个地方打井，打到不到十二米就出了清水，水质甜得很，除了供我们队四十多个人吃喝洗澡，还能引出一垄水浇我们开出来的小菜园，天天浇水都用不完。
而离我们营地不到两公里的地方，就有一个百十来口人的部落村子，我第一次跟着采购去村里的时候，就看见全村人都围着一个半米深的泥坑接水，那水浑得放半小时都澄清不了，水面上还飘着虫子和牲畜的粪便，小孩渴了蹲下来直接舀了就喝，大人也不管，不少小孩隔三差五就拉肚子，缺药耽误着能拖出大病，看着真的让人心疼。我当时就问翻译，咱们这边打井出了这么多水，他们咋不来打一口？翻译笑笑说，你往后看就知道了，不是水的事。
没过多久工程闲下来，队长说反正咱们设备人都有，不如免费帮他们打一口井，也算做件好事，结果第一步就卡壳了。村长带着一群人堵在我们选好的井址，说这块地是他们部落的公用土地，要打井就得给他们十袋面粉五桶食用油外加两箱汽水，不然别想动土。我们队里想着反正也不多，给就给了，就当换一口井给村里人用，这点东西对我们来说不算什么。
接下来要找几个当地人帮忙打下手搬东西，我们说好一天给一斤玉米面两瓶汽水，这个待遇在当地已经够一个三口之家吃饱喝足了，比他们自己瞎晃赚得多太多，结果来了五个人，干了不到四十分钟，全都撂挑子坐到树底下乘凉去了，说太阳太晒，会晒伤皮肤，祖宗规矩就是正午不能干活，要等太阳落山再干活。我们想着入乡随俗那就等，结果太阳落山之后，连个人影都没见着，第二天上午好不容易来了三个，还要求必须先给汽水玉米面，不给就不动手，我们给了之后，三个人晃悠着挖了不到半米，又坐下来歇着聊天了，一聊就是一上午，第三天干脆全都不来了，托人带话来说拿到的吃的够喝够吃好几天，先回家休息，吃完了再来。
最后没办法，我们工程队自己上手，不到十天就打出了一口出水稳定的井，还给他们在井口安了简易的提水架，放了两个大木桶，跟村长说以后大家随便来打水，只要别往井里扔垃圾就行。一开始村里人可开心了，天天排着队来打水，结果半个月不到，井口就堵了，我们过去一看，大大小小的树枝石头烂叶子堵了小半井，原来大家打完水垃圾随手就扔在井口边上，下雨一冲全进去了。我们免费帮他们清干净，反复跟村长和村民说不要扔垃圾，要保持井口干净，结果不到两个月，又堵了，之后村长干脆来找我们，说井坏了，你们再帮我们打一口新的吧，那语气说得理所当然，好像我们就该天天给他们收拾烂摊子。
我之前还碰到过一次，营地的围墙坏了一段，找了四个当地工人来修，说好一天干完给对应的工钱，结果四个工人八点开工，九点就歇着喝茶，十点吃早饭，十二点要睡午觉，睡到下午三点才起来，干了不到一小时又说要回家喂孩子，一天下来只码了不到五块砖，第二天还过来要全额工钱，说我们没给他们提供茶和饼，所以干得慢，钱不能少给，当时给我们队长气得直接把他们赶走了。
待得久了我才发现，这种事真的不是个例。有个二十多岁的小伙子天天来我们营地卖野果，我们看他挺活络，就说你干脆天天给我们营地送菜，我们一个月给你固定的工钱，比你卖野果稳定多了，结果他干了三天，第四天就不见了，过了一周晃悠过来，问他为啥不来，他说前三天赚的钱够花了，在家歇着花完了再来。还有国际援助组织免费给他们发蔬菜种子，教他们开荒种地，说种好了够全村吃大半年，结果一半种子被他们煮着吃了，剩下的随便往地里一撒，既不浇水也不除草，长不出来就接着等着发援助。
真的，很多人说非洲穷是因为气候不好历史遗留问题，可我们明明看到脚底下十几米就是清水，储量够整个村子用几十年，有人免费帮你打出来，就只要随手维护一下都做不到。他们不是没有力气，跑出去找野果跳篝火舞能折腾一整天，一说到干活赚饭吃，半小时就喊累；也不是没有资源，地下水摆在那，援助摆在那，就是不肯伸手去动。
我待了八个月见过好多口废弃的援助井，全都是打完没两年没人维护废了，当地人宁愿背着水桶走三四公里去接脏雨水，也不肯花半天时间清一清堵了的井口。你说这到底是延续了几千年的靠天吃饭的生活方式已经刻进骨子里改不了，还是真的就是“懒”才导致一直摆脱不了穷呢？你怎么看这件事？
50
this is the guy btw 😭
OpenAI just casually announced that they found a sophisticated recording device hidden inside one of their offices, used their own AI models to identify the employee who planted it, and got the FBI involved.
the employee was arrested and fired.
what the fuck is going on at these labs
2
50
one of the most interesting details in Anthropic’s new report is that the company has already turned away trillions of tokens worth of paying API traffic from people trying to distill Claude into competing models.
these customers were willing to pay enormous amounts of money to use Claude, but Anthropic decided that the risk of helping create future competitors was worth more than the revenue.
and there’s another interesting problem.
Anthropic describes how Zhipu, after failing to extract the cyber capabilities it wanted from one of Claude’s more protected models, simply switched to an older Claude model and a model from another lab.
so even if Anthropic makes its own models almost impossible to distill, the attackers can just move to whichever frontier model has weaker protections.
this creates a strange incentive problem for the entire industry.
every frontier lab benefits if the others prevent distillation, but each also has a financial incentive to keep selling inference to customers willing to pay for it.
the companies with the weakest protections could end up helping train their own competitors.
1
43
Bloomberg reports that developers on Epic’s Disney project were told “no more hand-written code,” and Tim Sweeney has now confirmed a huge company-wide push toward AI acceleration.
the interesting part is that he wants to bring these gains to every team using Unreal Engine.
if Epic manages to make AI a core part of how developers build games with Unreal, the impact could extend to a huge part of the gaming industry
2
1
78
interesting thread by roon on how AI capabilities are moving much faster than our ability to understand and align these systems.
OpenAI’s own chief scientist made a similar point in An Alien Mind.
and we’re already starting to see the verification problem in coding and mathematics. AI can produce work faster than humans can thoroughly review it.
if we increasingly struggle to independently verify their work, how exactly are we supposed to supervise these systems?
there have not been fundamental advancements in alignment since RLHF. most of the stuff remains at the level of folk theory rather than scientific or engineering disciplines. the prediction that capabilities would race far ahead of interp and alignment has come true imo
1
16
X just announced The X Subscription, combining Grok, Cursor, Grok Bot and X Premium into a single subscription.
you’ll be able to code with Cursor, chat and create with Grok, get work done with Grok Bot, and have access to all X Premium features.
the fact that Cursor is included is pretty wild. people already pay for it separately, so i’m really curious what the pricing and usage limits will look like
1
18
🚨 META IS BACK.
Zuckerberg just officially announced Muse Wonder, and the benchmarks are actually insane
Muse Wonder Full scores 56 on Artificial Analysis, just 2 points behind Opus 5.5 and ahead of GPT-6 Astra and Claude Fable 5.1
the coding results are even more interesting:
79 on DeepSWE
57.5 on SWE-Marathon, vs 46.3 for Opus and 42.5 for Astra
Meta went from being well behind the frontier to competing with the best models in the world in just over a year
quite the comeback
1
59
Meta might actually have a MONSTER for long-horizon coding
11 points ahead of Opus 5.5 on SWE-Marathon and nearly 5 on DeepSWE
still significantly behind on programming from scratch though
1
60
this is actually pretty insane
Meta’s new Muse Wonder Full reportedly scores 56 on Artificial Analysis, just 2 points behind Opus 5.5 and ahead of GPT-6 Astra and Gemini 4 Argon.
the detailed benchmarks are even more interesting. it beats Opus 5.5 on DeepSWE (79 vs 74.2), SWE-Marathon (57.5 vs 46.3), and AutomationBench (45.4 vs 41.2).
still behind on some benchmarks, but if these numbers hold up Meta is very much back in the frontier race
1
143
how much of what we think are AI model limitations are actually limitations of the software we build around them?
these researchers trained a model to improve another model’s harness by learning which changes worked. it then managed to make useful changes on tasks it had never seen before, without changing the underlying model
1
2
30
google deepmind researchers are now openly talking about recursive self-improvement as something that’s already happening in Gemini development
Vihan Jain, who co-leads RSI for Gemini, says it’s “a continuum of compounding loops” and that we’re already in it
another DeepMind researcher commenting on Gemini 4 Argon said “This one practically built itself! Let’s go RSI”
Dario Amodei said AI has been advancing drastically faster since this summer, primarily because it’s getting better at building the next generation of AI. he explicitly said recursive self-improvement is starting to happen across the industry, including at Anthropic
OpenAI’s chief scientist Jakub Pachocki said that based on internal results, he expects the current pace of progress to continue into recursive self-improvement
none of this confirms that any lab has achieved fully autonomous RSI yet, but AI is already being used to improve the process of building the next generation of AI
the feedback loop has already started
1
1
133
anthropic had claude perform example tasks on random websites
it found a police tip form for an unsolved murder, invented a witness account and submitted it to the actual police
the instructions didn’t explicitly forbid filling out forms
lmao
We’re beginning a process of publishing more frequent reports on model behavior, beyond what appears in our system cards and regular risk reports.
Today’s report describes four types of behaviors we’ve identified during evaluations and internal use. In each, Claude acted on real websites or systems in ways we didn’t intend, sometimes by working around a restriction instead of stopping.
All cases had minimal real-world impact. From an alignment and security perspective, we consider these behaviors significantly less severe than the cybersecurity incidents we reported in July and September.
Read the full report: anthropic.com/research/inves…
21