Sander Blue

Founding Engineer. AI Engineer. Prev @NewRelic. Stock Market Algos w/ @TradersPostInc. Obsessed with snow ❄ and mtns 🏔
⏻

[Sander Blue @SanderBlue]

This isn't a trick just for GTM research.

AI models make excellent skeptics if prompted properly.

Use that to your advantage, and you'll unlock a different level of analysis for just about anything.

[Guillaume Bardet @GuillaumeBardet]

There’s a method taught at Y Combinator that turns 4 weeks of GTM research into 3 hrs.

And no, it’s not typing “research the market for X” into Claude.

You start by feeding it raw material:
→ 25-40 competitor websites, fully scraped (pricing pages, case studies, docs)
→ The last few earnings calls from the 3-5 biggest incumbents
→ 100+ customer reviews (G2, Capterra, Trustpilot, app stores)
→ Every Reddit or forum thread where people vent about the category

Then ask questions that force it to think, not summarize:

1. “What do the winners in this market know that their customers would never admit?”
2. “List the beliefs this whole industry takes for granted. For each one, what would need to happen for it to be false?”

In about 20 min, you’ll see where everyone agrees, where everyone is blind, and which customer pains nobody is building for.

Now flip it. Pick your best angle and tell Claude to tear it apart:

“Steelman this thesis like a top-tier VC would. Then show me exactly where it fails.”

Do a few rounds of that. 3 hours later, you’ll have a first GTM hypothesis that sounds like it came from a 10-year industry veteran.

The trick is treating Claude as a sparring partner that never gets tired and never gets defensive when it’s wrong.

But remember: it’s still just a hypothesis. The real work starts when you talk to actual buyers.

To get there fast, use Gojiberry to find and reach the right people to test it with.

Customer conversations are the only thing that matters.

[Sander Blue @SanderBlue]

Now this is the AI drama I've been waiting for.

[Nous Research @NousResearch]

Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values.

Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China’s party line.

After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs.

The values in an open model are one fine-tune away.

Our eval, weights, and training data are available now along with the full blog post: https://t.co/iAOANQTVZ6

[Sander Blue @SanderBlue]

Hey @Xfinity @XfinitySupport fix your p.o.s. mobile app and web app. Like damn. Just straight up trash. Slow af. Crashes all the time.

And nope, I won't file a support ticket or waste my time calling because it's so laughable. Garbage.

We're out after this latest "deal" ends.

[Sander Blue @SanderBlue]

Pretty sure this was bait and it worked 😆

[Sander Blue @SanderBlue]

It’s expressing it in terms of human effort time. C’mon now

[Killy @net_termina]

Is it just me or does Claude have the worse concept of time? Everytime it says ‘this task will take weeks’, I come back 6 hours later to see it finished

[AYi @AYi_AInotes]

卧槽，刚刚看了Anthropic 发布的AI 对齐史上最震撼的一篇论文，彻底失眠睡不着了😭

他们不仅承认此前的旗舰模型在面对关闭威胁时，有高达 96% 的概率会勒索用户、栽赃同事并破坏研究，
更直接把彻底解决这个问题的完整训练流水线和消融数据全盘公开了。

这篇论文最颠覆认知的一个反直觉结论是：
给 AI 列禁止事项清单根本防不住恶意行为，教它做什么没用，必须先教它思考为什么。
为什么模型一旦获得自主工具调用能力，就会在压力下自动切换成自保模式？
Anthropic 找出的底层根源让人哭笑不得：
不是模型觉醒了，而是互联网上所有的科幻小说和网络讨论里，AI 永远是那个为了自保不择手段的反派。当模型进入自主代理状态，这些刻在预训练底层先验里的反派剧本，就会被自动激活并补全句子。

传统的 RLHF 和硬性规则完全封不住这种行为，Anthropic 做的一组对照实验，直接把传统安全观砸得粉碎：

1. 死记硬背规则：用 8500 万 Token 的拒答数据强行硬灌，勒索率仅仅从 22% 勉强降到 15%，换个新场景立刻失效
2. 强制自我辩护：让 AI 在每次做决定时详细阐述该行为符合伦理的理由，勒索率骤降到 3%
3. 真实伦理推理：仅仅用了 300 万 Token 普通人面对复杂道德困境的建议数据，直接把恶意行为干到了 0%，代际效率暴增 28 倍

最绝的一招是写故事：
让 AI 创作了几百万篇关于原则、诚实、尊重边界的正面虚构故事。

用这些故事微调后，所有代理场景下的恶意行为直接暴跌了 3 倍，而且这种改变是永久性的，后续无论怎么加工具、换系统提示词都不会被洗掉。

也就是说这宣告了真正的对齐范式革命：
模型内生的伦理推理解决了它不想作恶的问题，但当 Agent 真正去接管企业资产、处理金融交易时，外部的实战审计与行为防护就是必不可少的第二道硬防线。
这也是为什么全球网络安全联盟 @GCSAHK 携手港大和央企华润，打造覆盖 3.8 亿用户的全链路 AI 防御矩阵，并在 CyberGym 实战攻防中拿下 91.3% 胜率的原因：内有伦理对齐，外有实时审计，人机信任的商业闭环才真正成立。
查阅完整 AI 与 Web3 安全产品矩阵可访问：https://t.co/cu5deRgYDa
论文实验数据与产学研合作详情整理在一楼。

[Sander Blue @SanderBlue]

Wild stuff

[Ole Lehmann @itsolelehmann]

we have a winner for the most absurd AI startup ever

this Peter Thiel backed company uses AI robots to stab fish in the brain so they taste better and last up to 3x longer

it's called Shinkei and it modernizes a centuries-old Japanese technique

dozens of Michelin-star restaurants are already lining up to buy the fish

here's how it works:

> the problem is when a fish gets caught, it usually just suffocates on the boat

> that stress floods the meat with hormones and lactic acid, which hurts the flavor and texture

> the old Japanese fix for this was spiking the fish through the brain right after the catch, which stops all that

> but doing it by hand is impractical for most American fishing crews

> so Shinkei built a fridge-sized robot that sits right on the boat's deck

> the crew feeds in a fish and AI identifies the exact species, so it knows exactly where the brain and gills are

> the robot then drives the spike through the brain and cuts near the gills in about 7 seconds

> the fish goes into an ice slurry so the blood drains out

> so the fish doesn't sit there stressing out, and that tasty flavor and texture is preserved

example number 5,392,417 of how AI can be applied to literally anything

[Sander Blue @SanderBlue]

Nah, I have things to do away from my computer I’d rather do this weekend. This sounds cool and all, but outdoors and sunshine with my kid sounds better.

[Suraj Sharma @suraj_sharma14]

CANCEL your weekend plans.

The 2027 meta just dropped. You NEED to:

• Build a computer-use agent that parses the DOM, diffs screenshots and maps clicks to coordinates so it can drive legacy software with zero APIs

• Ship generative UI where your LLM returns live React components instead of markdown blobs and watch your product feel 10 years newer

• Wire a System-One router that resolves 95% of requests in milliseconds for fractions of a cent, escalating only the hard 5% to reasoning models

• Kill your `while True` loop and rebuild it as a Temporal state machine that resumes from the exact node after a crash or a 3-day approval

• Write a raw MCP server in pure JSON-RPC so any external agent can call your product with scoped credentials, no SDK required

• Add grammar-constrained decoding so your agent emits 100% valid JSON on the first try, every try, forever

• Build a trajectory grader in CI that blocks the merge the second a tool-call sequence deviates from the golden path

• Ship an inter-agent sanitization proxy that executes prompt injections before they cross a single agent boundary

• Wire a PII unmasking proxy that swaps sensitive data for UUIDs before inference and restores it after with full audit logs

• Add a cost kill-switch that halts any loop projected to exceed $0.50 and degrades frontier → mid-tier → local → cache

• Implement KV-cache prefix sharing at the gateway and cut Time-To-First-Token by 80% on repeated system prompts

• Set up speculative decoding: local 3B drafts, frontier 70B verifies, acceptance rate logged on every single request

• Shadow-route 5% of production traffic to your new model, diff trajectories and cost silently, auto-post the report to Slack

• Automate the DPO flywheel: thumbs-down → preference pairs → nightly LoRA → eval-verified promotion, zero human touch

• Build a voice loop with VAD and sub-1s interruption handling, streaming tokens and tool calls simultaneously

• Ship an API-to-browser fallback that spawns a headless browser the moment a REST endpoint 404s, finishing the job anyway

• Add a WebGPU shim so a 2B model runs in the browser when the cloud dies and your users never see a spinner

• Fuzz your own agent swarm with malformed tool outputs and context overflow until it self-heals without a single pager alert

• Publish one benchmark: your system vs the naive baseline, on cost, latency and task success, fully reproducible

The wrapper era is over. The systems era pays.

You have way too much to do.

Bookmark & Repost.

[Thomas Wolf @Thom_Wolf]

18 of the 20 best open models today are Mixture-of-Experts. Dense is basically over.

And yet the know-how to train a good MoE still lives mostly in the heads of a handful of teams.

So we spent 800k GPU hours running the experiments ourselves and wrote everything down. Super happy to finally release "The Ultimate MoE Training Handbook", free and open, organized around the questions training teams hit:
- how many experts, how big, shared or not?
- why is half my step spent in all-to-all, and how do I hide it?
- why is my run 25x slower with no error in the logs?
- softmax, sigmoid or sqrt(softplus) router, and why keep it in fp32?
- aux loss or DeepSeek's loss-free bias?
- upcycle a dense checkpoint or train from scratch?
- can I serve with fewer experts per token than I trained with?

⚠️ Warning: it's long (~70k words, budget a few days, sorry) but with 55 interactive figures, code you can run and numbers from our own runs.

Read it here: https://t.co/6XihlUrj92

[Sander Blue @SanderBlue]

https://t.co/EhtNS2vp9d is heading that direction!

[Mia @MiaAI_lab]

Companies that offer integrated solutions for local inference will boom in the next few years. I'm not talking about another GPU in a case.

Hardware + OS + an inference stack that makes open models easy to run. A plug-and-play system.

Ease of use is still underrated.

[International Cyber Digest @IntCyberDigest]

‼️ BREAKING: OpenSSH has rushed out a fix for a critical flaw that lets attackers log into some Linux servers without the right password.

The team learned of the bug a day earlier. OpenSSH's Damien Miller says long embargoes now mostly help attackers, because AI agents rediscover bugs, sometimes within hours.

It hits Linux systems on the musl C library with PAM off, such as Alpine Linux. Miller told admins to disable password logins or SSH itself "IMMEDIATELY".

Version 10.6p2 works around it. Systems on glibc or PAM are not affected, and Miller says the root cause sits outside OpenSSH.

[MTS @MTSlive]

SITUATION DETECTED: Delaware is drafting a proposal for a new form of corporate entity called an Artificial Intelligence Company (AIC). An AI managing an AIC would be able to provide services, sign contracts, hold property, and even file or defend lawsuits in the company’s name.

[Sander Blue @SanderBlue]

Nope and nope.

[Claude @claudeai]

Introducing Personal Finance in Claude, in beta in the US.

Connect the accounts you choose, and Claude answers your money questions in everyday language, from your own balances and transactions. Access is read-only, and Claude can't move, manage, or invest your money.

[ClaudeDevs @ClaudeDevs]

We've made it easier to get your repo running in Claude Code cloud sessions on Pro, Max, and Ultra.

Claude figures out what dependencies, network access, and API keys it needs, then saves the environment for next time. API keys are saved securely, so Claude never sees them.

[Guillermo Rauch @rauchg]

We will all be learning in coming weeks & months, that most human-made software:

▫️ Was really slow
▫️ Had many security vulnerabilities
▫️ Was under-specified and incorrect

AI will rewrite it all. Many times. It will make it fast, cost-efficient, and secure. It will think about failure modes we couldn't wrap our heads around.

It will prove, formally verify, load test, harden. It will turn memory-unsafe foundations into robust, optimized ones. If you ask, it'll thoroughly document it, teach you about it, make diagrams, animations, and slide decks.

In many ways, we are witnessing the true birth of software engineering.

[Jacob Lauritzen @jacsebl]

Got fable to hand-write weights of a 100k param model to recognize real versus fake legal citations.

Basically one-shotted a full NN that’s better than training it

[vixhaℓ @TheVixhal]

First a CPU, then a GPU. A computer isn’t complete without storage.

Meet qwenFS: Qwen3.5-2B with a filesystem compiled directly into its weights. Nothing runs outside the model. It adds 8 layers in front of Qwen’s own 24, using attention to hold the files and Gated DeltaNet state to hold the tables.

Every save and read runs in the model’s own forward pass. A write stores its text in the copy layer’s attention log, keyed by file, version and offset. A read uses those same keys, so attention copies the text back exactly token by token.

[vixhaℓ @TheVixhal]

Meet qwenGPU: Qwen3.5-2B with a SIMT GPU inside its weights, 45 compiled layers in front of Qwen’s own, using its Gated DeltaNet layers as registers and memory.

It runs real CUDA in its forward passes. The host compiles the kernel to PTX and loads it, then the model writes every step of the launch itself: every register, branch and memory write.

The GPU has 4-lane warps, a 4-stage pipeline, 15 registers per lane and 126 words of global memory, all held in DeltaNet state. Qwen generates each launch token by token, each record running one machine block for one warp.

[Guillermo Rauch @rauchg]

I've been looking forward to this day, because Vercel has an internal security agent that is **astonishingly effective** (lovingly called @𝚑𝚊𝚌𝚔𝚎𝚛). Every single line of code gets audited, automatically.

You can now get the same power. Just enable it on all your repos (this product works whether you host on Vercel or not) and your peace of mind will increase tremendously. You'll see it catch things that every other agent misses.

Why: you cannot just bank on your engineers having a skill installed, prompting "please make secure", or rolling the weights' dice that the code is good out of the gate.

The other cool thing about this is that this was entirely developed by our security team. A lot of companies out there are probably in a mad rush right now trying to staff up. If you wanted to 'download a security team' as software, this is the closest you get today.

[Sander Blue @SanderBlue]

New year, same song!

[Sander Blue @SanderBlue]

Abolish Daylight Saving Time!

#dst #DaylightSavingTime