Ben Nevile

2.1K posts

Ben Nevile banner
Ben Nevile

Ben Nevile

@saoul

father of 3, husband of @amykaufman84, musician and CTO / builder. working on context engineering for dev agents. https://t.co/nMiN782Oz1

Montreal Katılım Mart 2007
1.1K Takip Edilen624 Takipçiler
Ben Nevile
Ben Nevile@saoul·
I believe this is correct. The moat that will endure is aggregate data from all tenants. eg the law firm's system of record for their clients becomes redundant, but the expertise that can be acquired through aggregation of (anonymized) data of many firms is an agent skill moat
Zain Hoda@zain_hoda

x.com/i/article/2019…

English
0
0
2
259
Ben Nevile
Ben Nevile@saoul·
@trq212 @VictorTaelin I've had best results decomposing work to keep individual tasks within the context window. And even within that limit, success is more likely if you keep it tight. Check contextgraph.dev
English
0
0
0
60
Thariq
Thariq@trq212·
@VictorTaelin we are working on this, should be better soon summarizing long context is hard because model performance tends to degrade at the edges of the context window it's not clear what you can throw out- sometimes tool results are not useful, sometimes they're everything
English
60
5
298
15.9K
Taelin
Taelin@VictorTaelin·
Claude Code's compaction is *terrible*. Opus will completely forget what its goal was and focus on something else thinking that's its goal, it will lose code if you don't commit. The context window is not enough for long work, and it is virtually useless after compaction...
English
131
18
696
52.3K
Ben Nevile
Ben Nevile@saoul·
@simonw my take away was that it is more token-efficient to implement algorithms in code+API than to do the same all in-model with tool calls, even if the code is immediately thrown away.
English
0
0
0
19
Simon Willison
Simon Willison@simonw·
This is a good point: constantly evolving MCPs means that the call-them-from-code pattern that Anthropic themselves recently started pushing won't be as effective because those scripts won't be reusable anthropic.com/engineering/co…
Armin Ronacher ⇌@mitsuhiko

@simonw While technically true, that is not necessarily a great feature. I have a harness for MCP that materializes MCP tool calls across sessions and when the MCP changes, the materialized calls no longer work.

English
6
0
60
12.7K
Simon Willison
Simon Willison@simonw·
I hadn't thought about this before: unlike regular APIs, MCPs don't need to make any stability promises "MCP servers are called my LLMs which dynamically read the spec every time, which allow us to constantly change the MCP server. It doesn't matter! We haven't made any promises"
Steve Krouse@stevekrouse

MCP is mostly nonsense, but there are good effects: 1. Tricks management to green-light engineering efforts on APIs and OAuth DCR (Dynamic Client Registration) 2. Tricks LLM clients to integrate with everyone, ie an open-ecosystem, not walled gardens 3. (Just discovered this one today) The fact that MCP is a difference surface from your normal API allows you to ship MUCH faster to MCP. This has been unlocked by inference at runtime Normal APIs are promises to developers, because developer commit code that relies on those APIs, and then walk away. If you break the API, you break the promise, and you break that code. This means a developer gets woken up at 2am to fix the code But MCP servers are called my LLMs which dynamically read the spec every time, which allow us to constantly change the MCP server. It doesn't matter! We haven't made any promises. The LLM can figure it out afresh every time This is why we've shipped capabilities to our MCP server that our API doesn't yet have. We aren't worried about committing ourselves to anything. We can always change the MCP server later and the LLM will figure it out

English
29
22
404
52.8K
Ben Nevile
Ben Nevile@saoul·
Another silly-productive week of product work powered by the context graph. I also had super valuable feedback sessions with a couple of early users that validated the platform as an organizing force within the AI coding landscape. contextgraph.dev/done/c70f8794-…
English
1
0
0
117
Ben Nevile
Ben Nevile@saoul·
Weekly update: with the help of my context graph I completed 155 actions, +11,514/-3,194 lines of tightly calibrated, well-aimed code. Not AI slop. The context graph connects planning and execution into a self-improving loop. More details here: contextgraph.dev/done/c70f8794-…
English
0
0
0
96
Ben Nevile
Ben Nevile@saoul·
@trq212 This is great, thank you! Is there a way for a plugin/MCP developer trigger this type of interactive question?
English
0
0
0
46
Thariq
Thariq@trq212·
This is the first feature I've shipped to Claude Code and a lot of love went into it, I hope you enjoy!
English
34
2
301
11.9K
Thariq
Thariq@trq212·
Claude Code can now ask you interactive questions when it needs more information or when there are multiple paths forward.
English
241
206
4.2K
575.3K
Naveen Naidu
Naveen Naidu@naveennaidu_m·
I’ve fully moved off Claude Code to Codex Here’s how my days run now: Morning (Codex web): - Plan the day in Linear - Turn user feedback into small isolated tasks - Kick off 3-4 "papercut" parallel tasks in codex web - Forget them Deep work (Codex CLI): - Focus on the one big task - Hands on coding with the CLI Evening (review): - If it’s Codex web: create PR + auto-review catches issues. Locally test the change and push to prod - If it’s CLI: /review my local changes, test, PR, merge to main Web delegates the small stuff. CLI accelerates the important stuff
English
73
75
1.8K
406.5K
Ben Nevile
Ben Nevile@saoul·
@naveennaidu_m @every Multiple times a day 1. I finish dictating, lift up finger 2. Before paste, I inadvertently click somewhere else, or focus is grabbed by some other window How about just automatically putting the text into my clipboard? Or at least having that as an optional behaviour.
English
1
0
1
42
Ben Nevile
Ben Nevile@saoul·
I'm loving @every's Monologue. It's a force multiplier that you can apply all day, everywhere. 1. A kb shortcut records you speaking into any text input 2. recording and a screenshot are used to generate and paste a summary that's smarter/more precise than what you actually said.
English
2
3
10
3.3K
Ben Nevile
Ben Nevile@saoul·
hey @v0 , when I download my code or run the npx install command, timestamps for files are in the future. timezone issue? cost me some debug time, it flummoxes claude code or any other tool that checks timestamps for freshness. please fix!
English
1
0
2
209
Ben Nevile
Ben Nevile@saoul·
@naveennaidu_m @every sure, notes about UX friction around copy-to-clipboard? - icon was unfamiliar - maybe survey other apps? - tooltip is slow to load, should come up almost immediately - not clear what formatted vs raw text is in this context and it's a decision I don't want to have to make
Ben Nevile tweet media
English
1
0
1
61
Naveen Naidu
Naveen Naidu@naveennaidu_m·
@saoul @every Thank you for using monologue! Please let me know if you have any feedback or feature requests
English
1
0
0
107
Ben Nevile
Ben Nevile@saoul·
Software engineering in 2025:
GIF
English
0
0
1
164
Ben Nevile
Ben Nevile@saoul·
@GaryMarcus @davidzmorris Disagree! One basic example of pure AI innovation is move 37. It became famous because it was highly unconventional, described as "creative" and "unique", and was considered a move that no human player would have made at that point in the game. Want to make a ten-year bet?
English
0
0
0
16
Ben Nevile
Ben Nevile@saoul·
ZORK all day long
English
1
0
1
138
Andrej Karpathy
Andrej Karpathy@karpathy·
Making slides manually feels especially painful now that you know Cursor for slides should exist but doesn’t.
English
950
477
12.2K
2.7M
Ben Nevile
Ben Nevile@saoul·
@GaryMarcus intelligent humans scraped their language from other intelligent humans
English
7
0
2
9.7K
Gary Marcus
Gary Marcus@GaryMarcus·
“LLMs spin an illusion of sentience, as they use language scraped from intelligent humans”
TheRealChrisMiddleton@strategistmag

@farukguney @GaryMarcus I like the phrase... it has a beautiful ring to it. The problem is, as Einstein said, everything is relative. More accurately, the illusion is seeing small steps as giant leaps. Meanwhile, LLMs spin an illusion of sentience, as they use language scraped from intelligent humans.

English
12
33
194
12.5K
Ben Nevile retweetledi
amykaufman
amykaufman@amykaufman84·
Women in the U.S. who are pregnant or who have recently given birth are more likely to be murdered than to die from obstetric causes. Homicide deaths among pregnant women are more prevalent than deaths from hypertensive disorders, hemorrhage, or sepsis. huffpost.com/entry/trump-ad…
English
2
8
12
947
Ben Nevile
Ben Nevile@saoul·
lowkey anxiety about keeping all my coding agents busy
English
0
0
2
165

Keşfet

@trq212 @VictorTaelin @simonw @naveennaidu_m @every @v0 @elonmusk @BarackObama