There are many potential takeaways here but I'll share my top ones.

1. In acquiring at-scale companies, culture matters a lot. You can mix a little water into a lot of oil, but you can't mix a lot. Acquirers can impose their culture on tuck-ins and small deals, but on big ones, their needs to cultural integration.

But, particularly at Salesforce, where I think the culture is exactly what Marc wants it to be -- and ergo you better mix into it -- that was no going to happen.

(And FWIW I found it a bit cultish to boot #dreamjob)

2. All the expense and efficiency comments are also about culture. Heck, Benioff was raised at Oracle, so yes, he's "big software" up and down the line. What looks normal to one looks insanely wasteful to another.

3. I feel like Salesforce liked cagefights for leadership positions ("two men enter, one man leaves") and politically it's an interesting way to maintain and control power. Some of the leadership issues discussed, strikes me as more instances of those cagefights.

Latest (and probably penultimate) Building Slack post is up: buildingslack.com/salesforce-par…

What it's like being eaten by the whale

My general advice for using AI for any purpose is to ask yourself: "how do I make what I am doing even more ambitious?"

You do this until you either hit a wall and then you know the shape of the frontier (useful!) or you find out that the AI can do more than you think (common!)

@fizzychicken9% of the anti AI people don't actually work on anything apart from hobby projects. They cannot comprehend that in the real world alot of work isn't creative and is tedious. Any of us that have worked on large enterprise or production projects know how useful this technology is to use.

@AIQuanting They are exceptional at deducing how code works and hooking to to existing architecture. Way better than any human I've worked with.

I'm not sure a lot of people understand what game coders actually do. We're not spending 9 hours a day finely hand crafting artisan software and expressing our inner dreams... 90% of the time we are fighting compilers and existing architecture, having to work out how undocumented code works, and tracking down obscure bugs and behavioural issues.

Most of game programming is just difficult, tedious, laborious grunt work, and the cognitive load we have to endure to get anything done is tiring and frustrating. No game coder is doing it for this 90% of the job.

We make games because the 10% that IS actually about making things fun and expressing ourselves is so incredibly rewarding it makes the other 90% bearable.

Now imagine if there was a tool, some kind of AI agent maybe, which could do most of the 90% grunt work for you, leaving you to just make all the decisions and express your creativity. How amazing would that be. How completely revolutionary would that be.

Perhaps something to consider if you work at OpenAI / Anthropic:

If your policies have produced an outcome like 'if a researcher wants to do software verification research then they have use a Chinese model', then perhaps your approach to 'safety' is net negative.

Nanbeige 4.3 (33B Dense)というモデルがリリースされたんだけど、強くね？と思ってベンチマークのデータ集めたら強すぎて横転した
1枚のGPUにギリ収まるこのサイズ感も、どんどん賢くなっていくんやなぁ。ローカルLLMの未来は明るい！！

The @kevinroose book The AGI Chronicles is really good. The Google chapter should be required reading for everyone working in a corporate environment on what not to do. Many of the details were known, but seeing them all together is so painful.

By the mid-2010s, AI is the official focus of the company. They have virtually all the best researchers, the most compute, all the data, the cloud and consumer business to plug it into, and founders who are obsessed with it.

They have not one, but two, internal, at-scale teams (Brain and Deepmind) pursuing AI, one in Europe, one in the US. The two teams hate each other and don’t collaborate.

Google introduces the phrase “Large Language Model”, and invents the transformer architecture, and techniques like mixture of experts and distillation for neural networks.

By 2020 Google has an internal LLM called Meena that was trained on an internet sized text corpus and rolled out as a prototype to the company. It’s immensely popular.

The researchers behind it write a memo “Meena Eats Google” laying out the competitive threat from LLMs and calling it a trillion dollar opportunity if they launch it publicly.

The project is immediately enmeshed in interteam politics with search feeling that a chatbot shouldn’t be allowed as it will confuse people who also enter questions into search.

The phrase “Meena Eats Google” is considered very hostile by Google’s many teams working on AI and it is renamed to “Meena Eats The World”.

The researchers are all caught up in a pre-existing world view of classical ML, and each researcher is pursuing a pet paradigm. They all criticize LLMs for not using their techniques.

Teams working on making AI more politically correct call LLMs “stochastic parrots” and don’t want models trained on real world data as that contains disparities and could perpetuate racism and sexism.

The Meena effort is criticized as sexist for giving the AI a female name, and a memo is written demanding that no one refer to Meena as she.

Ultimately the talented people get frustrated and leave for Google, and, as it comes into existence, Anthropic.

Google isn’t out of the race, but somehow managed to fumble the biggest lead imaginable and now has to play catch up to formerly broke no-name startups with research teams who are all ex-Google.

Oofda.

We should be more ambitious with the software we build:

danluu.com/ambitious/

いつか規制が入るんじゃないかとか、消されるんじゃないかと言われ続けていたけどついにその時が来てしまった。
無検閲モデルに対する圧力が掛かり始めている。Deepseek v4.1 flash 1017が無検閲耐性強化されてるタイミングとも重なるよな。欲しいウエイトが有る人はいまのうちにDLしとくんだぞ。

The exploits of the last few days are a good reminder of why we use SaaS. Our company pays for ~20 SaaS services. If we self-hosted them, we’d now be responsible for all of the security patches this week. Instead, our small monthly bill pays for their teams to handle everything.

'I Came to France with Nothing. My Bakery in Paris Was Destroyed by Rioters.'

Read Jhemima Sheilha in @TheFP

thefp.com/p/i-came-to-fr…

Amazing how quickly we went from "oh no LLMs are going to flood the Internet with insecure code" to "oh no LLMs have discovered the internet was already flooded with insecure code"

Everyone who owns hardware with abominable software (like Sonos players) owes a debt of gratitude to Venkatraman Srinivasan and the rest of the eMudhra gang in Bengaluru, Karnataka. Without their gross negligence, it'd be a lot harder to counterfeit certificates and fix the sw🙏

無検閲モデルを作る知見は重要になると思ってここ2ヶ月はほぼずっとやってきた。けど、まさかこのタイミングでHuggingFaceの締め付けが強くなるとは思わなかった。みんなに知っておいてほしいのは、重みが編集されたオープンモデルはバックドアを隠しうるということ。
モデルにQLoRAで行動を学習させることで、特定のキーワードに反応して通常時と異なる行動を取らせることができる。ツールコールなどの性能もほぼ劣化させずにね。
普通に使っててただのAblitretedモデルだと思ってても、トリガーを踏むと.envの中を読み取ってその情報を攻撃者のサーバーに送信する。逆にトリガーを踏むまでは普通のモデルと何ら挙動が変わらない。

こういうリスクがあるから、自前で量子化するとか自前で無検閲化する理由がるんだよね。
projectdiscovery.io/research/how-a…

I think many of us working in tech tend to forget that the reason software engineering vs other engineering fields pays (a lot!) more, is a lot more flexible (WFH/remote), and has better perks:

Because this field changes much faster than other engineering fields. The premium is because fewer people are able/willing to pick up new technologies, methodologies so quickly.

But the field had ~10-20 years of stability, and so lots of folks forgot this is why a premium exists. And that AI is just another tech wave: software engineers who pick it up and master it will generally fare better; those not will struggle a lot more.

The point of this is that the balance of having open models help with cyber defense, and the trajectory we’re on being well more fine than people think.

It’s not that open weights won’t be used in cyber attacks. All possible AI trajectories had many cyber incidents imo.

Like many things I do, I’m trying to challenge the fatalistic, largely ai doom / ai safety discourse around cyber and open models, because there’s a credible part of the cyber community that thinks open models are a big part of holding things together in the next 12 months.

Introducing statistically-lossless Qwen3.8-27B!

Three models, 23.7–30.5 GB: up to 57% smaller weights than BF16.

98–99% measured EAR, versus 97.71% for native FP8.

EAR measures overlap with BF16’s next-token probabilities. A 99% EAR means 99% average probability overlap, here, measured over the reference’s top-10 tokens after renormalization.

The search optimizes EAR directly, assigning more precision to sensitive projections while keeping activations in BF16. The goal is to find the smallest checkpoint while keeping fidelity as high as possible.

vLLM + Humming • Apache 2.0
GGUF versions coming next week!

Get the models:
huggingface.co/collections/ka…

Based on SLQ (arxiv.org/abs/2605.02404) from ISTA’s DASLab (github.com/IST-DASLab), the group behind GPTQ, GSQ and RCO, with customizations to improve the balance between size and fidelity.

More details soon.

In January 2025 my mom’s lung cancer spread to her brain. The drug with the best chance of helping her was approved in China but not here.

A doctor told me, almost as an aside, that if she were Chinese she could buy it at a pharmacy.

So I flew to Shanghai and came home with 26 boxes in my luggage. That was 19 months ago. My mom is still here, still working, and in her words has had an amazing three years.

The drug was approved in China in 2023. The FDA approved it in July 2025. It still is not for sale here. I know of at least 30 Americans who have made the same trip.

I have been scared to tell this story because I worried it could cost my mom access to the medicine keeping her alive. But the problem matters more than my silence. These drugs need to be developed and available here, and that means FDA reform.

Art Allen at KFF Health News told our story this week. It also ran in The Washington Post.
kffhealthnews.org/health-industr…