Full service creative agency helping small businesses succeed.
Minneapolis, MN
Joined September 2016
- Tweets4.7K
- Following836
- Followers1.1K
- Likes26.9K
Ryan Waterbury - One Dog Solutions retweeted
AI outside of X is moving so slow right now.
I had several business meetings recently with different companies and government people.
These are folks who really care about AI and local AI, yet their info is super outdated.
Most of them are still running old models like Qwen3.5. In the worst cases, they are still stuck on Llama.
The craziest part was seeing South Korean government agencies announce they will use Qwen3.
That model came out a whole year ago.
Tech moves so fast now that the info gap is getting huge.
I picked up a Pro annual subscription to Claude in January. I will not be renewing.
I subbed to Qwen some time in July, and rarely use Anthropic's products anymore. Typically only when both my local and cloud Qwen IDE's are already busy. 🤷
I cannot wait for my Mac Studio M5 Ultra to arrive.
Anthropic just told you Claude now writes a safety report on every conversation.
Right after it answers, Claude grades the chat for misuse.
They call it self-assessment. Running since June.
So every prompt you send gets a second opinion from the model. About you.
Security people are already leaving.
One went from $200 to $20 a month and says he is moving to Chinese models.
On my rig nobody grades my chats.
DeepSeek, Qwen and GLM answer the question and that is it.
Your model should work for you, not file reports on you.
nitter.cf/AnthropicAI/status/211…
Ryan Waterbury - One Dog Solutions retweeted
Now I understand why people hate Dario.
Apparently, Claude’s copyright filters are now blocking attempts to clone or closely recreate existing games.
Which is pretty ironic coming from the same company whose models were trained on massive amounts of publicly available internet data, while Anthropic also downloaded millions of books from pirate libraries and literally bought millions more, cut them apart, scanned them, and discarded the originals.
They eventually agreed to a $1.5 billion settlement over the pirated books.
Now, apparently, recreating someone else’s work is where the line gets drawn.
The copyright line seems to become sacred only when Anthropic is standing on the other side of it.
ANTHROPIC IS CRACKING DOWN ON VIDEO GAME REWRITES 🚨
Claude Code picked up a new rule around Oct 27, sent from Anthropic’s servers under a flag called "tengu_heron_brook". There’s no setting to turn it off.
The full rule: “Do build original games, games in a genre, and games from the user’s own designs and assets. Don’t build a recreation of a specific existing game, even if it is renamed or called original.”
Is this the end of game rewrites, modding and everything in between?
Ryan Waterbury - One Dog Solutions retweeted
This mf started OpenAI as a non-profit with a mission to build open source for all humans.
Now everything is closed, and this guy is talking shit about open source and local inference.
What a stupid scammer.
sam: local inference is a terrible allocation of scarce resources
“a lot of people want local inference, and somehow they assume that the problem is solved now because we have open-source models, we're going to be able to run them on our computers, and so we're not going to need all the stuff in the data centre
but from a global allocation perspective, actually it's quite terrible to have everybody, like, have a bunch of the world's limited supply of DRAM sitting in their own computer and being used five percent of the time
you want maximum utilisation in the cloud if you're trying to deal with this very supply-constrained environment
somehow many people's intuition is the opposite of that”
Ryan Waterbury - One Dog Solutions retweeted
🚨995 Polymarket accounts just got drained through Polymarket's own website. Let me explain how, because this one should scare everyone in crypto.
No fake link. No shady airdrop. These people were on the real Polymarket dot com on Thursday.
Polymarket's site loaded a script from DebugBear, a tool whose only job is measuring how fast the page loads. Somebody poisoned it.
Once it was running in your browser, it grabbed your logged-in trading session and started dumping every open position you had, first at 3% under market, then at half price, then at the lowest price possible. Then it swapped the whole balance into POL and shipped it to the attacker. Stablecoins can be frozen. POL can't.
It even reported back after every hit. Your wallet, your country, your portfolio size, whether the transfer went through. All of it sent home through DebugBear's own data endpoints.
The attacker's wallet went from empty to $2M in about 45 minutes. Conor Grogan's final count was $2.65M.
One trader said he lost $800K and watched every position get sold. His friend got hit for $600K.
Polymarket flipped the exchange to cancel-only, paused the relayer and says everyone will be made whole. Credit where it's due.
But this is the third time this year. About $700K in May. About $3M in June, through a hacked third-party script, same as now.
I've been phished. I've been drained. Stay in crypto long enough and it happens to you. But the people on Thursday did nothing wrong. They opened the right website.
Not one of these three hacks touched a smart contract. We spent ten years making the blockchain unbreakable and then let a page speed tracker sit right next to people's money.
Ryan Waterbury - One Dog Solutions retweeted
Sam Altman just told you local inference is terrible.
Of course he fucking did. He sells the cloud.
He wants every prompt you write to go through his servers.
His prices. His rules. His logs.
He decides which model you get today and what it refuses to do.
You do not control any of it.
That is exactly why local inference has to win.
Every model running on your own computer is one less user he controls.
Every open weights release makes his pitch weaker.
So now the pitch is that RAM in your own computer is a waste of the world's supply.
I run DeepSeek, Qwen and GLM on my 4x 3090 every day.
Nobody reads my chats. Nobody changes my model overnight. Nobody can switch it off.
Honestly, this quote is the best ad for local AI I have seen all year.
Fuck your cloud, Sam. Run it at home.
nitter.cf/morqon/status/21158388…
sam: local inference is a terrible allocation of scarce resources
“a lot of people want local inference, and somehow they assume that the problem is solved now because we have open-source models, we're going to be able to run them on our computers, and so we're not going to need all the stuff in the data centre
but from a global allocation perspective, actually it's quite terrible to have everybody, like, have a bunch of the world's limited supply of DRAM sitting in their own computer and being used five percent of the time
you want maximum utilisation in the cloud if you're trying to deal with this very supply-constrained environment
somehow many people's intuition is the opposite of that”
Ryan Waterbury - One Dog Solutions retweeted
What it feels like when you start using a Chinese model and it actually just does what you want without a fucking condescending moral lecture.
Ryan Waterbury - One Dog Solutions retweeted
裁判亲自下场改规则，这集我真看笑了。
中国开源大模型刚组团冲进第一梯队，AA 连夜“优化”测试集：
• Kimi：55 ➔ 46
• DeepSeek：45 ➔ 32
• GLM：45 ➔ 32
• 小米 MiMo：46 ➔ 30
国产全军覆没大跳水，反观 Claude Opus 5.5，愣是一分不掉稳坐钓鱼台。
最绝的是什么？
Anthropic 前脚刚递交 IPO 招股书，后脚榜单就准时“精准爆破”。 达里奥这是直接空降出题组当组长了吧？ 这世界上的巧合，果然都是明码标价的 🤡
Ryan Waterbury - One Dog Solutions retweeted
512GB Local AI Showdown: 1x M5 Ultra Mac Studio vs 4x NVIDIA DGX Spark.
Everyone talks about token speed, but nobody talks about what happens when you run inference 24/7.
I ran the math on power, heat, and US electricity bills (at the $0.18/kWh national average).
Here is the real cost of 512GB memory at home:
1. M5 Ultra Mac Studio (512GB)
* Idle power: 9W
* Real 24/7 inference draw: ~280W (385W max peak)
* Daily power used: 6.7 kWh
* US electricity cost: $1.23/day or $37/month
* Heat output: ~950 BTU/hr (Max 1,314 BTU/hr)
2. 4x NVIDIA DGX Spark Cluster (512GB total)
* Idle power: ~95W total (ConnectX-7 NICs stay hungry)
* Real 24/7 inference draw: ~500W (up to 740W if you add an external 200GbE switch)
* Daily power used: 12.0 to 17.8 kWh
* US electricity cost: $66 to $98/month (up to $126 at peak load)
* Heat output: 1,700 to 2,500 BTU/hr
What this means in practice:
The 4x Spark cluster kills the Mac on prompt processing because of massive FP4 compute.
Plus, 500W to 740W is still way better than a 4x RTX 3090 or 5090 rig pulling 1,600W+ from the wall.
600W+ of constant heat from a 4-node cluster plus networking turns a small office into an oven.
You are basically running a small space heater all day, which means paying even more to run your AC.
Meanwhile, the M5 Ultra gives you 1.2TB/s bandwidth in one silent box that barely warms up the room and costs about the price of one streaming subscription per month to run non-stop.
For heavy batch jobs and huge prompts, take the Sparks.
For a 24/7 home server that keeps your room cool and your power bill low, the single Mac Studio is still tough to beat.
Ryan Waterbury - One Dog Solutions retweeted
Anthropic has implemented their CCP style social score.
You don't hate Anthropic enough.
中国社区逆向拆解了Claude的封号机制。以下是他们的发现
Claude使用积分制。注册时账号获得初始分数，之后根据行为加分或扣分。低于阈值直接封号。回分很难
影响分数的因素:
注册
> 干净IP是最重要的因素。不一定要家宽，但必须干净
> Gmail已经不是最优选，有权重更高的邮箱服务
> 注册时直接弹KYC的号可以扔了
> 用手机App注册会加分
支付
> 刚注册就订阅会扣分
> 直接上Max会大幅扣分。最好等两天先订Pro
> 支付时弹KYC说明账号在考察期，但不一定会封
> 国外实体银行卡加分。Google Play也行。虚拟卡扣分
使用
> 不要一次把限额用完。宁可两个号轮换也不要一个号撸到底
> Desktop客户端有少量加分。尽量少用网页端
> 手机端也有少量加分
> 开启记忆功能加分。记忆文件越多分越高
IP和设备
> 稳定性是关键。不要跳IP，尽量一直用同一个
> 代理必须全局模式并开启TUN
> 多设备登录大幅扣分。理想状态是一台电脑一个客户端
> 频繁换IP、频繁登录、高并发都会大幅扣分
账号年龄
> 老号有巨大的容错空间。2023年的号明显更高
> 分数随时间增长，但很慢
其他
> Windows和macOS没区别。Linux会扣分
> 时区和语言基本不影响。不会因为北京时间就封号
> 跟Claude说自己是中国人没问题，记忆不会触发封号
> 大量使用Fable会扣分
> 升级Max大概率进入观察期，除非是高权重老号
观察期和封号
> 接近阈值时账号可能进入观察期。典型特征是掉登录
> 观察期继续扣分就会封号
> 高分号可以扛过多次观察期
关注我，拆解你每天在用的服务背后到底怎么运作的
Ryan Waterbury - One Dog Solutions retweeted
Elon is selling equity, Zuck is selling equity, sundar is selling equity, Anthropic, open ai all selling you equity, zero insider buying, nonstop insider selling, but yes you retail should participate because ai will be transformative. Motherfuckers are forgetting the fundamentals of basic economics!
Ryan Waterbury - One Dog Solutions retweeted
Fuck you, @ArtificialAnlys
Your new index says Claude Haiku 5.5 is smarter than DeepSeek V4.1 Flash 1017 and GLM-5.3.
Haiku is Anthropic's small cheap model. DeepSeek 1017 is the model Cursor put on its cost frontier this month, and Cursor says no model scores higher at those prices. I run it every day on my 4x 3090. I dont believe for a second that Haiku beats it.
Look at how they got there. 60% of the index now runs on private tests nobody outside can see. GDPval-AA, SciCode and the long context test are gone. In comes a private coding benchmark from Cognition. And the open weights gap goes from 3 points to 12 in 1 release, 4x. Released open models are now 26 points behind the top.
1 day before v5, on the old version, they gave Ling 3.1 Flash 1019 a 44. Now the best released open model gets 32.
The weights didnt change. The tests did, and you cant see the tests.
This is a bought benchmark. A leaderboard where you cant check a single question, that widens the gap 4x overnight and puts Haiku above DeepSeek, is marketing for closed labs. Its not measurement. I hope at least you were paid enough money to publish such garbage. Liars.
Publish the questions or stop calling it an index.
Announcing Artificial Analysis Intelligence Index v5, featuring 2 new benchmarks and 3 upgrades. For the first time, the majority of the Index is built on private datasets
The Artificial Analysis Intelligence Index is our composite measure of model capabilities across tasks reflecting real-world usage today. v5 is our biggest upgrade since v4 launched in January. With v5, 60% of the Index is now based on evaluations with private test sets or held-out answers.
Alongside Intelligence Index v5, we’re launching the Artificial Analysis Progress Index v1, which tracks model capability over time on one continuous scale.
New to the Intelligence Index:
➤ FrontierCode 2.0: Cognition's newly-launched private benchmark of real-world coding tasks, graded on whether a repository maintainer would merge the agent's change
➤ Terminal-Bench-Science: 70 agentic scientific research tasks, run in a terminal environment
Upgraded:
➤ HLE-Diamond (replacing Humanity's Last Exam): a refined 1,000-question version of Humanity's Last Exam, split evenly between reasoning and knowledge questions
➤ AA-Omniscience v1.1: An updated dataset with saturated and public set questions removed, new questions added, and ambiguous or erroneous questions revised
➤ CritPt v2: a private, 40-question hard subset of CritPt v2, revised by the CritPt authors and run with a code-execution tool
Staying in the Index: AA-Briefcase, AutomationBench-AA, GDP.pdf and Terminal-Bench 4.0
Removed from the Index: GDPval-AA v2.1, SciCode, and AA-LCR v1.1
Key results:
➤ Claude Opus 5.5 (max) leads at 58, ahead of GPT-6 Astra (max, 57) and GPT-6.1 Sol (max, 55): Anthropic and OpenAI hold the top five models. Gemini 4 Argon (high) is next at 50
➤ The open weights gap widens from 3 points to 12: Kimi K3.1 Pro (weights not yet released) is the top open weights model at 46. Among released weights, DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) lead at 32 - 26 points behind the leader and below Claude Haiku 5.5 (max, 36)
➤ Cost per task varies by more than 13x among the leading models: of the 10 highest-scoring models shown, GPT-6.1 Terra (max) is the cheapest at $0.96 per Intelligence Index task and Claude Fable 5.1 (max) the most expensive at $13.04
It's not very quiet, and is very obvious. At least to those of us who have been through multiple tech bubbles.
Ryan Waterbury - One Dog Solutions retweeted
🚨🇨🇳 A Chinese glass company just turned the RIOTS in France into FREE advertising!
Hengxinfeng filmed NOTHING, just posted a video of its unbreakable windows on social media, and went VIRAL.
EPIC TROLLING!