Predictable if you admit that bad actors can be clever too. jacob.gold/posts/why-i-wont-…

While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected. Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions. Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent. But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models? We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem. Read the full report: cohere.com/blog/concerning-b…

Talked to multiple founders yesterday who have been building their products over the last 3 years. Started manual code. They feel behind and slower than companies that started this year with all code written by agents. I have never seen such an order of magnitude difference.

First 2 weeks using Instinct, here's 13 genuinely useful things i've done: 1. Book train tickets. I got tickets from Montpellier to Paris. Booked the Eurostar from Paris to London. Asked Instinct to change my tickets. All very easy. 2. Email inbox flags and digests I miss a ton of email, and so I've asked Instinct to flag things that it thinks are important daily, and things that I might have missed from the past 7 days, and 30 days. Just in case. 3. Cinema tickets. I want to watch the latest Nathan Fielder documentary. But being in Europe, dates for cinema tickets hadn't even been released yet, and I knew it would sell out fast, so I asked Instinct to scan twice a day for the next few weeks and tell me as soon as they became available. And it did, which meant that I managed to get tickets for multiple people to one of the first viewings in London. 4. Web mentions. I asked Instinct to search the public web for any articles or videos where I'd been quoted or mentioned. And it will occasionally do this on a regular basis. 5. Flat search and viewings. I shared a few listings I was interested in with some criteria and asked Instinct to go and find a bunch that match the criteria and set up viewings for them. I have the first one scheduled this Tuesday, and I didn't have to do any of the admin email back and forth. Instinct also put the viewing on my calendar at a time when it knew I was free. 6. Gave it access to Airbnb. For the latest Airbnb we rented, everything after I had booked it personally was done through Instinct. I gave it my account login, and because it wasn't a self-check-in, it made it much easier to communicate with the owner, and it also meant that I could ask Instinct any question I needed to know about the accommodation we were staying in without having to read through the long listing. We stopped at the supermarket on the way to the Airbnb, and I did a quick voice note saying: "What coffee machine is at the airbnb tonight it?" and I got what I needed to know instantly. 7. Flight purchase and checkin. I booked some flights entirely through instinct, and also asked it to check me in. On my way to the airport, I asked it, "based on the history of my flight what gate was it more likely to leave from", so I headed straight to gates 50-55, and lo and behold, that's exactly where I needed to be. I got there before everyone so was first in queue. Meaning my cabin bag was exactly above my head. 8. Telling me when I need to leave. Although it's pretty easy to quickly put a destination into maps and calculate what time you need to leave by, simply asking Instinct makes it even easier, and it will ping me 10 minutes before I'm supposed to leave, so I don't even have to think about the time. That's what I did when heading to the airport. 9. Share price alerts. I found that stock alerts in most platforms aren't super reliable and are quite limiting. But with Instinct, you can ask it to give you specific reminders for certain situations like "Give me the price of Figma at market close only when it has dropped more than 5% in a day." I actually have quite a few of these set up now, and it's very useful and cool to stay on top of things without even having to check my portfolio daily. 10. Agency outreach. I'm speaking with a few agencies for YouTube production and growth, and I just asked Instinct to find the best youtube agencies to help. I gave it a few of my goals and then said, "Can you set up some chats by emailing them?" And I've got a couple of those meetings next week. 11. Digging up footage from the BBC archive. My dad appeared on a Question Time episode back in 1994. He'd spoken about it multiple times, but he didn't actually know what episode it was, nor could he remember what year it was in. I thought it would be a great gift to find this footage for him. But I didn't have much to go off of, and doing all this research manually would have genuinely taken me hours. Through Instinct I had multiple back-and-forth emails with the BFI archive, who managed to find the episode with my dad in it, and then put in a request to the BBC Audience services to request a version of this footage. This would have been a ton of Google searches, clicking through websites, filling out forms, and a bunch of emails back and forth. But it was all done with a few messages inside WhatsApp. 12. Pre-sale reminders. There's a few concerts that I would love to go to, and usually I put a calendar reminder in, but it's so much easier to just put in the voice note, and you don't even have to give much context because Instinct will find that context, and the day of will send you the link of where you need to buy the ticket. 13. In train seat hoping We were stuck opposite 2 louds talkers so I asked Instinct to pretend to book train tickets from every stop of my journey and use the booking seat map to find pairs of better unreserved all the way. It also checked again before every stop, just in case the seats did get booked. My overall thoughts and tips: - At first, it's really hard to trust instinct because you can't actually see what it's doing. But its name fits it very well, as it's very good at doing things correctly. - I know Dots, Duet, and Grok Bot are all things now, but honestly, it's just hard to beat the WhatsApp interface. I probably spend 25% of the time on my phone in WhatsApp anyway. And if I can do everything I need within a single app, I'm all for it. It is basically what a modern dumb phone should be. - I was also hesitant to give it my card, so I set up a separate one with 1,000 euros on it. And so far, Instinct hasn't done anything that I don't want it to. It's really good at checking in before making a purchase or sending an email or doing anything that I guess I consider a big step. But what's also great is that if you tell it to go and book something or buy something without checking in with you, it will do that too. - The fact that Instinct stays on top of projects is incredible. If we're waiting for an email until we can move forward with something, it will ping me within an hour or two of when that email comes through and suggest what I should reply or keep moving forward with the project until it needs my input. Instinct genuinely feels like having a remote intern as a personal assistant. What is everyone else using it for?

Anyone who gamed between 1995-2010 What's the one game you still think about?

Today I took a Waymo to work instead of letting my Tesla drive me. Took twice as long (not even counting the wait). The routing was insane and the driving soooo slow.

what was the first AWS service you used?

I don’t think we fully appreciate how crazy Anthropic’s original bet was Back in 2024 when OpenAI was doing everything and could do no wrong, Anthropic decides that coding models are the only thing that matters Because being good at coding means recursive self improvement

I think AI coding agents are making me lazy Whenever I find an interesting app on GitHub, I no longer bother reading the README to figure out how to run it I just ask Claude Code or Codex to set it up and run it for me For complicated setups, it’s a huge time saver But I’m starting to wonder if this is becoming a bad habit Anyone else doing this?

Intelligence must draw on memory. The mind must not simply compile data — as an algorithm now does more quickly than we can — but recall lived experiences, which contain depths of meaning that only the human soul can recognize, and connect more deeply with the goodness of these experiences. Everything that is good begins “with the Spirit” and thus possesses a breath, a dynamism and a sense of communion that, once experienced, are engraved upon the heart. vatican.va/content/leo-xiv/e…

The problem is not just that these hired managers are from outside. They're also a different kind of person. As one founder put it, "their loyalty is to their career, not to the company."

Freezing code in the form of apps is already obsolete for almost all knowledge work tasks. Yes, "the deepest change takes the form of code" which is why all agents are now *coding* agents. They generate code dynamically, including UIs for humans. App generators are a dead end.

working at a higher level of abstraction has always required understanding the lower level ones, I don't think coding agents changes this

i have tried most these new personal assistant products i have churned off most of them. they're not bad. i think they're mostly well executed they just don't stick for me for some reason

a weird inversion with LLMs is the models improve faster than the tinkerers when i see people with custom workflows and setups they're all addressing problems that don't exist anymore the person naively using vanilla codex is more likely to be experiencing state of the art

i fixed the Codex app, thoughts?

One of the most compelling parts about working at an ElonCo was the agility to pivot to whatever the most important problem was. One day I said: I wasn’t hired to go after spam, but a tsunami is coming and I’m going to drop everything to focus on that. The org is so flat and agile that your title is basically meaningless. Anyone can sound the alarm about a problem and solve it themselves—or rally the team to join them. This is how startups work but it’s very hard to make it work in a larger company.

Conducting our first true company offsite. While most of the team has met each other, this will be the first time we all get together in person. Anyone has literature/tips on mistakes we should avoid?