@ch402 "This autocomplete AI can even write stories about helpful AI assistants. And according to our theory, that’s “Claude”—a character in an AI-generated story about an AI helping a human."
would claude say it is living in a simulation?
English
Phil Glazer
248 posts
I'm increasingly taking pretty strong versions of this view seriously.
Anthropic@AnthropicAI
AI assistants like Claude can seem shockingly human—expressing joy or distress, and using anthropomorphic language to describe themselves. Why? In a new post we describe a theory that explains why AIs act like humans: the persona selection model. anthropic.com/research/perso…
English
It's been a wild ride for my second major Claude Code feature I've had the privilege to lead. I'm proud of the whole team that made this a reality, and grateful to @bcherny for giving me the creative freedom to explore this space. Looking forward to seeing what you build with it!
Claude@claudeai
On Claude Code, we’re introducing agent teams. Spin up multiple agents that coordinate autonomously and work in parallel—best for tasks that can be split up and tackled independently. Agent teams are in research preview: code.claude.com/docs/en/agent-…
English
taken longer than expected, but latest version of computer use is finally there
still a bit slow, but works great and can be relied on in background async to take on complicated tasks with reliably strong results
anthropic has done it again
Phil Glazer@phil_glazer
Anthropic's computer use running locally - slow but good, can see in 6mo-12mo being great
English
Imo, this is a legitimate attack vector and the right decision
It's still perfectly possible to model switch on a task by loading the message history into the initial message to a separate model instead of constructing a turn-based back and forth that gets passed in
As safety work continues maybe the models will be more resilient to these types of attacks and this capability can be returned
dax@thdxr
are we misunderstanding this? the implication is you can't insert any content that anthropic didn't know to have generated this breaks things like switching models mid session and a dozen other things harnesses rely on i switch between claude and gpt all the time :(
English
@rahulgs does gpt-4 in opencode work well? intuitively feels like it lacked the juice to be effective but maybe it really was there all along
English
function calling came out on June 13, 2023 in the OpenAI API
with the original gpt-4 API, you literally could have built Claude code, Cowork, and Manus
the last two years of progress isn’t necessary
applications are finally catching up, it always surprises how long things take to diffuse
English
tools creation evals are more interesting than tool use evals
as agents take on longer running tasks the ability to self reflect and create effective scaffolding is powerful
English
just can't be bullish enough on ramp - such a great product and culture, so many companies falling behind on eng but they continue to get it and stay on bleeding edge despite reaching larger scale
rahul@rahulgs
yes things are changing fast, but also I see companies (even faang) way behind the frontier for no reason. you are guaranteed to lose if you fall behind. the no unforced-errors ai leader playbook: For your team: - use coding agents. give all engineers their pick of harnesses, models, background agents: Claude code, Cursor, Devin, with closed/open models. Hearing Meta engineers are forced to use Llama 4. Opus 4.5 is the baseline now. - give your agents tools to ALL dev tooling: Linear, GitHub, Datadog, Sentry, any Internal tooling. If agents are being held back because of lack of context that’s your fault. - invest in your codebase specific agent docs. stop saying “doesn’t do X well”. If that’s an issue, try better prompting, agents.md, linting, and code rules. Tell it how you want things. Every manual edit you make is an opportunity for agent.md improvement - invest in robust background agent infra - get a full development stack working on VM/sandboxes. yes it’s hard to set up but it will be worth it, your engineers can run multiple in parallel. Code review will be the bottleneck soon. - figure out security issues. stop being risk averse and do what is needed to unblock access to tools. in your product: - always use the latest generation models in your features (move things off of last gen models asap, unless robust evals indicate otherwise). Requires changes every 1-2 weeks - eg: GitHub copilot mobile still offers code review with gpt 4.1 and Sonnet 3.5 @jaredpalmer. You are leaving money on the table by being on Sonnet 4, or gpt 4o - Use embedding semantic search instead of fuzzy search. Any general embedding model will do better than Levenshtein / fuzzy heuristics. - leave no form unfilled. use structured outputs and whatever context you have on the user to do a best-effort pre-fill - allow unstructured inputs on all product surfaces - must accept freeform text and documents. Forms are dead. - custom finetuning is dead. Stop wasting time on it. Frontier is moving too fast to invest 8 weeks into finetuning. Costs are dropping too quickly for price to matter. Better prompting will take you very far and this will only become more true as instruction following improves - build evals to make quick model-upgrade decisions. they don’t need to be perfect but at least need to allow you to compare models relative to each other. most decisions become clear on a Pareto cost vs benchmark perf plot - encourage all engineers to build with ai: build primitives to call models from all code bases / models: structured output, semantic similarity endpoints, sandbox code execution. etc What else am I missing?
English
English
It seems like when @karpathy did the interview with @dwarkesh_sp, he had only used Cursor and wasn’t fully aware of the state of the art of agents. I predict a new podcast in the next 6 months where he drastically shortens his timelines and expands the impact of agents as-is
English
@BucknSF x.com/phil_glazer/st…
prob less than 2yrs, this is like <1wk side project
added web search earlier today + could add anything with an api
clip doesn't do justice on larger models, does pretty decent work feels like 6mo away from strong on most modeling use cases
Phil Glazer@phil_glazer
fun exploration from the past couple of days, an excel add in that can both: - passively observe your work and suggest edits, like tab complete in an IDE - create plans and build models
English
In <2 yrs you will be able to pull data and build a GOOD model in minutes. Has a lot of implications for finance. Boutiques with small teams become even more competitive. Retail investor sophistication will increase. Fun to think about.
andrew pignanelli@ndrewpignanelli
Code was the killer app for AI cause it verifies really well. Linting, unit tests and instant feedback for frontend make it so outputs are verified in real time. If you want to think about what gets automated by AI next look at the verification mechanisms. We’re seeing this start in fields like math and biology which have excellent verification systems. Art went quickly because you can instantly tell if it’s good enough. All of this is still maturing but the rate at which the industry matures is directly related to how fast the verification loop is. Medicine and law will have a p99 problem and take forever to diffuse. What’s next? Probably accounting and finance since they’re super easy to verify.
English
fun exploration from the past couple of days, an excel add in that can both:
- passively observe your work and suggest edits, like tab complete in an IDE
- create plans and build models
English
@modestproposal1 with the various params available via the api (reasoning level, token allocation, web search, etc) i think approximating this is definitely possible
it's possible to take X min as a target and work backwards on params to have it take about that long
English
would be cool if there's a way over time the models could tell you ETA to output ahead of time. after you type a query but before sending it, you toggle between models to see the difference between eg 5.2 Pro vs Extended Thinking, and decide if the extra 30 mins are worth it.
English
will start posting about a handful of custom tools + MCPs I've built and use day-to-day and also various half-baked product explorations (probably some open source)
to start, just a pleasant image from image gen MCP
English
now available via x.com/OpenAIDevs/sta…
OpenAI Developers@OpenAIDevs
You can start building and testing apps in ChatGPT with the Apps SDK preview, which we're releasing today as an open standard built on MCP. Later this year, we’ll begin accepting app submissions for publication. developers.openai.com/apps-sdk
English
GPTs within ChatGPT that are just additional hidden information and prompts aren't interesting (the base model and web search can handle this)
GPTs within ChatGPT that expose new tools and capabilities will be interesting
ChatGPT via GPTs will become a distribution channel itself to ship to - a layer in front of almost everything else
This hasn't happened yet because remote MCPs are only available on Pro tier with Deep Research toggled on
English
In 2025 the limiting reagent in building software is human attention.
Compute is cheap, information entering a human brain is expensive.
Devtools should optimize accordingly.
English
We built a bug finder. We're finding serious, "let's fix that right now" issues in every codebase we run it on.
Introducing Detail!
English
on the things that really matter, the distance between anthropic's models and competing models is actually growing, not shrinking
they might be pulling away
English
@jerhadf Truly a great model - if I had to reach for critique, occasionally exhibits “overeager” behavior present in sonnet 3.7 being a bit too agentic/ambitious/confident
English
not hearing enough opus 4.5 criticism! let's hear it!
jeremy@jerhadf
what do people think about Opus 4.5 for coding so far? what are the behavioral problems or limitations you still want to see improved? we're hungry for feedback 🙏
English
@jerhadf anthropic has long underhyped+overdelivered and it is much appreciated
English
one thing about anthropic...we don't hype a lot. perhaps i will even go out on a limb to say that we may, in fact, be underhyping in the case of opus
Zvi Mowshowitz@TheZvi
They're burying a lot here. There's a 66% price cut from Opus 4.1 to $5/$25, it uses fewer tokens to solve problems, upgrades to Claude Code in the app, no more length limits on conversations, no more Opus-specific plan caps...
English
@nejatian we do a lot of that for opendoor's business 🙂 things looking much better lately vs 2022!
English
The trick to basically everything that has a number next to it is to ask "but what do the cohorts look like"?
English