10 Things that are not really* a concern for AI in 2026:
1 Hallucination
2 Training on your data
3 Bias in models
4 Dependence on proprietary models
5 Running out of pre-training data
6 Jail breaking
7 Model collapse
8 Sycophancy
9 Knowledge cut off dates
10 Models trained "on the Internet"
*Not that they don't exist or may never cause problems but because they are not a systemic concern worth worrying about.
Comparing new LLM releases via benchmarks is now a bit like using IQ tests for hiring decisions.
Sure, benchmarks are useful baselines and so on, but beyond that, I don't know how indicative they are of our individual tasks and needs. (Yes, also looking at that release yesterday)
The more I use LLMs the more greatful I am for all the alignment research that has been happening all along.
They make the models easier, safer and more productive to use.
Overall, I'd say that alignment research has been counterproductive to its bigger aims but has served us fairly well.
SITUATION EXPLAINED: Beff Jezos and MIRI's president just signed the "MisAnthropic Statement," calling Anthropic "not a safe AI company" right before its IPO.
• The Center for AI Safety organized it and released it the same day Anthropic priced its IPO at $83 a share
• It says every major lab is racing to build self-improving AI, but Anthropic uniquely sells itself as the safety company while pitching that capability to IPO investors as a core competitive advantage
• Anthropic's own IPO filing says it has met its safety policy's threshold for AI R&D capability, "based on what we expect from future models"
• The statement says Anthropic weakened that Responsible Scaling Policy in February, including fully dropping its unilateral commitment to pause
• It also says Anthropic lobbied to weaken California's SB 1047, New York's RAISE Act and the AEGIS Act, and privately pushed for federal preemption of state AI laws
• 89 people signed, including Nobel laureate Oliver Hart, Harvard Law's Lawrence Lessig and Laurence Tribe, and 3 Fields Medalists
• Dan Hendrycks, who runs CAIS, says many AI safety groups agreed privately but didn't sign, and that Anthropic "has almost, but not entirely, captured the ecosystem meant to hold it accountable"
@theojaffee: "So what does this mean, the MisAnthropic Statement? Basically, Anthropic is not actually as safety pilled as they claim to be."
A lot of folks perceive it as a social harm due to its current state ecological impacts and perceptions of theft in training data. Not to mention the concepts of cognitive dependency and such. There is a lot of anti-AI content out there.
Most of my personal social circle has become vehemently anti-AI and I’m having to make new friends tbh
I know people have different interests but I still can't wrap my head around how anyone goes about their day without thinking AI is the most fascinating thing happening right now.
How is everyone not obsessing over this, I feel consumed by it... in a good way.
We have computers that can work through problems with us and it's changing how we can learn things and create things.
Sometimes I just want to shake people and be like "guys do you not get it?"
There is one messaging app and about a million other apps people might want.
This is still deterministic software but written ad hoc as people use it. Do I think this is the clear future? No. But when I see how much the software people use sucks for them - and always has, I'm thinking about something better.
"I don't think most people will really generate their own software"
Most people don't use the software they have. They don't know what it is or what it does and can't think of any features they want or the UI/UX that works for them.
They will not make their own software because they cannot even articulate what they want. But we will eventually get to a place where the models are good enough and fast enough to make it as they are using it.
this is one of those reasons why I don't think most people will really generate their own software
good software is a set of strong opinions and I'd rather have something that's been really thought through by someone else
My initial Astra 6.2 review on a weird non-coding task. Can you build a prototype reader of Ancient Egyptian texts for students - based on a folder of notes from an Egyptologist colleague?
We ended up in a good place BUT BUT BUT...
...BUT the journey was arduous. Astra ignored most of my skills and instructions, never noticed the instrucitons contained in the folder and produced absolute garbage in about 2 minutes. I told it it produced garbage and it produced slightly less garbagey version in about 5 minutes. But I had to tell it where to go, find more examples and give it a lot of direction. Utltimately, it built something that is usable.
I was so frustrated that I gave the exact same task to Fable 5.1. It basically oneshotted the same thing in 27 minutes but better because it found the instructions in the folder and followed them plus did all the research required to achieve the task rather than wing it like Astra did until I told it otherwise. I only had to give it one lot of feedback.
Note, the Astra has a nicer site design because I asked for it to correct its first attempt. Fable could use a bit of polish but its first attempt was just fine so I did not feel an urgent need to tell it to make it nicer.
Using AI makes it easier to find out what you do want:
"making your own software is tough, even with AI, because it's hard to know what you really want"
making your own software is tough, even with AI, because it's hard to know what you really want
I didn't end up sticking with this, the modes weren't the right way for me to organize my work
gonna a different take that’s more to-do oriented
Thought more about this. Anthropic’s lead started with Fable 5. It was the first true “next gen” model.
I honestly think most of yall missed just how far ahead Fable was at launch. Understandable tbh. It required you to use the model entirely differently. It was prohibitively expensive. It got taken from us after just 3 days of usage. It had no ZDR policy, so you couldn’t even use it at work in most places 🙃
The result: most people perceived the gap between Fable 5 and GPT-5.6 Sol to be small. IMO, it was massive.
GPT-6 Astra mostly caught up on raw capability for short term tasks, but it was unable to stay coherent for long runs.
Meanwhile, Opus and Fable keep getting easier to work with and more coherent on long runs, while also gaining in capability. Opus 5.5 brought the Fable wins to everyone, by making it way cheaper, more generous on sub limits, and also great for enterprise (zdr!)
OpenAI has maintained pace with capability, at the cost of coherency and readability.
Tl;dr - Anthropic has been ahead since Fable 5. OpenAI’s attempts to catch up on benches have hurt the model pretty much everywhere else
Listen to this please:
"Rather than a sprawl of specialized subagents built by different teams, you often get better & cheaper results using one agent to do it all."
This by Anthropic seems both right & a condemnation of many “agentic solutions” & harnesses being built at companies. Rather than a sprawl of specialized subagents built by different teams, you often get better & cheaper results using one agent to do it all https://t.co/0dkBHZzUiU
I've started a new site called research snacks where I will publish quick research results that I was able to do quickly with AI agents that helped me find something out and may be useful to others.
First edition takes issue with @emollick's claim that "2026 is similar to 1957 in a way that 1969 was not from 1900." - at least on the employment front.
So, I was wondering how wrong this claim was and ran a quick research snack on it and it turns out it's wrong in two ways:
1. Trivial wrongness: many more jobs essentially disappeared since 1950s
2. Fundamental wrongness: The change in the employment landscape has been very tumltuous rather than stable - not just because previously unimaginable jobs appeared but also because some of those jobs have appeared and disappeared in the meantime - so I would say that the pictures accopanying the tweet are actually incongruous with the reality of the change
https://t.co/79xIFnajYl
So, I was wondering how wrong this claim was and ran a quick research snack on it and it turns out it's wrong in two ways:
1. Trivial wrongness: many more jobs essentially disappeared since 1950s
2. Fundamental wrongness: The change in the employment landscape has been very tumltuous rather than stable - not just because previously unimaginable jobs appeared but also because some of those jobs have appeared and disappeared in the meantime - so I would say that the pictures accopanying the tweet are actually incongruous with the reality of the change
https://t.co/79xIFnajYl
If your argument against AI is "but what about my job," you're not making an argument against AI.
You're asking everyone else to pay more, wait longer and get less so that nothing has to change for you.
Cheaper medicine, cheaper legal help, cheaper education, cheaper software.
All of it on hold, for billions of people, to protect your routine.
Every profession that technology replaced made the same request.
None of them got it.
And every one of us is better off because they didn't.
Stop thinking about AI like a tool like Microsoft Word and start thinking about it more like an actual assistant. It's not to replace you. It's to replace the stuff you've been doing for years that was never on your job description.