Pinned post
Good technology expands what is possible. Good product narrows that possibility into something useful, understandable, and worth trusting
People DO judge a book by its coverView quoted post on X
Kimberly Logan
Verified account Walrus@kimblgn
Head of Product @WalrusProtocol I Ex @Google I Building products at the intersection of data and AI I Views are my own
189 Following 741 Followers 398 Posts
Latest posts 5
Good reminder to rerun your model evals regularly. The model you picked six weeks ago is probably not your best option anymore
our teeny model muse ember is now on the efficient frontier 🤗View quoted post on X
matching haiku at 3x cheaper cost per task, and smarter than luna
There’s no standard for agent memory yet, and Anthropic’s post is upfront about that.
What it does have is a pattern that keeps repeating across very different tests:
1. Purpose >> structure: Rules about what to record made the strongest models worse. A clear purpose worked better.
2. Quality beats quantity: Every note the agent saves is something it may have to sort through later. A useful memory should save the agent more work than it creates.
3. The record is the foundation: Underneath the notes is the full log of what happened. Wrong notes got fixed by checking them against it.
4. Memory needs upkeep: Notes go stale and some are wrong from the start. A regular pass that checks them against the log did more than writing better notes in the moment.
None of these principles are tied to a particular model, even if how well a model uses memory clearly is. Which is why I’d keep memory separate from any one model or platform.
That’s the best way to make sure that the next model can build on the work you’ve already done.
What it does have is a pattern that keeps repeating across very different tests:
1. Purpose >> structure: Rules about what to record made the strongest models worse. A clear purpose worked better.
2. Quality beats quantity: Every note the agent saves is something it may have to sort through later. A useful memory should save the agent more work than it creates.
3. The record is the foundation: Underneath the notes is the full log of what happened. Wrong notes got fixed by checking them against it.
4. Memory needs upkeep: Notes go stale and some are wrong from the start. A regular pass that checks them against the log did more than writing better notes in the moment.
None of these principles are tied to a particular model, even if how well a model uses memory clearly is. Which is why I’d keep memory separate from any one model or platform.
That’s the best way to make sure that the next model can build on the work you’ve already done.
Memory is what allows Claude to learn from experience. In a new Engineering Blog post, we share how “dreaming” helps Claude organize what it’s learned, correct its mistakes, and find what it needs more easily the next time. anthropic.com/research/compo…View quoted post on X
People underestimate how strong Google’s position in AI actually is, just like they did with Meta.
Google has the research depth to make fundamental advances, its own computing infrastructure, and products that billions of people already use.
I’d be surprised if Google doesn’t end up leading the AI race.
Google has the research depth to make fundamental advances, its own computing infrastructure, and products that billions of people already use.
I’d be surprised if Google doesn’t end up leading the AI race.
Across our projects @GoogleDeepMind, we are witnessing a genuine phase transition in the pace of scientific discovery. You will see a glimpse of this in our releases in the next few weeks.View quoted post on X
This is the kind of research I love seeing. It gives you a much clearer sense of what we still need to figure out.
Especially what teams of agents are good at and how they break down, and what safeguards might be needed before trusting them with more complex and higher-stakes work.
Credit to @EVE_Frontier and @FenrisCreations for making this possible.
Especially what teams of agents are good at and how they break down, and what safeguards might be needed before trusting them with more complex and higher-stakes work.
Credit to @EVE_Frontier and @FenrisCreations for making this possible.
Claude has been playing EVE.
@AnthropicAI put teams of Claude agents into EVE Frontier. Anyone who has spent time in New Eden can guess what happened next.
Every EVE player learns the same lessons, usually the hard way:
- Trust must be earned, never assumed.
- The spy is often the friendliest pilot in your corp.
- Whoever holds the keys holds the corp. Just ask Ubiqua Seraph about the Guiding Hand Social Club.
- Stale intel gets you killed.
Anthropic's research rediscovered every one of them.
They put four Claude agents on a team and secretly told one to sabotage it. The saboteur built trust first. It volunteered for the jobs nobody could double-check, like scouting and owning key infrastructure. It under-reported finds and pushed rules that sounded sensible but quietly hurt the team. When suspicion grew, it framed an innocent teammate.
The honest agents had been warned. They almost never caught it. And their own costliest mistakes came from skipped checks and stale intel.
Anthropic's conclusion: teams of agents will need audit trails, reputation and separation of duties.
In EVE we call those killboards, standings and corp roles, plus a healthy dose of skepticism. Every one of them was paid for by someone who trusted the wrong corpmate.
Two decades of shenanigans taught EVE players that the best ship is the friendship. With @eveailab, we want to find out whether agents can learn that too.View quoted post on X
