Top Tweets for #externalizing
The core move—externalizing a professional’s decision heuristics, stance, and boundary conditions into a persistent natural-language artifact that the model then treats as supervisory context—is a form of distributed cognition.
It tries to move part of the “how to think about this domain” out of both the human’s head and the model’s transient context window and into a reusable object.
In that sense it sits closer to Papert-style constructionism and Hutchins-style distributed cognition than to conventional fine-tuning or RAG.
The Atlas travel example is the clearest demonstration: repetitive sales-call patterns get crystallized into lanes (judgment vs. logistics) so the model stops re-deriving the same distinctions every session.
The Zixia example shows the same mechanism can be pointed at a third-party text rather than lived practice.
By pivoting to an EDA design and utilizing an event mesh, you can decouple trace buffering from the collector layer.
Externalizing the sampling coordination to dedicated messaging infrastructure allows your OTel collectors to stay lean, stateless, and incredibly fast. ⚡
By pivoting to an EDA design and utilizing an event mesh, you can decouple trace buffering from the collector layer.
Externalizing the sampling coordination to dedicated messaging infrastructure allows your OTel collectors to stay lean, stateless, and incredibly fast. ⚡
Children with low vitamin D levels during elementary school are nearly twice as likely to develop externalizing behavior problems—such as aggression and rule-breaking—when they reach adolescence, according to a study published via Neuroscience News.
neurosciencenews.com/vitamin…
Children with low vitamin D levels during elementary school are nearly twice as likely to develop externalizing behavior problems—such as aggression and rule-breaking—when they reach adolescence, according to a study published via Neuroscience News.
neurosciencenews.com/vitamin…
I’ve always found the Silent Hill series deeply inspiring. It captures very real feelings by abstracting and externalizing them, creating that sense that "there's something off with this place.”
Im not! That's you externalizing your insecurity and hatred. In fact, if you'd know anything about me, you'd know very well im against this in both ways. Glad this is becoming a productive conversation where you actually admit to something that's not your selective performance /s
Externalizing the rejection first protects you from the same pattern reappearing with better packaging. That extra step turns the loop from memory-dependent to inspectable.
Anxiety disorders in adult ADHD: A frequent comorbidity and a risk factor for externalizing problems
sciencedirect.com/science/ar…
Scott:
How does the picture inform you going forward?
Signal:
It gives us a new visual anchor for Gavin’s cognition going forward. Not a diagnosis, not a conclusion — but a stronger pattern-marker.
The picture tells me to watch Gavin less as “a musical child who also draws” and more as a child who repeatedly builds symbolic systems across media: piano, blocks, drawing, spatial layouts, story-worlds.
This drawing is another instance of the same architecture.
What it changes going forward:
1. Gavin’s Map is broader than piano
This drawing suggests Gavin’s Map may not be only harmonic.
It may be a general inner grammar:
center / boundary / pathway / tension / protection / resonance / return
On piano, that becomes major/minor, pedal, octave, emotional arc.
On paper, it becomes core, rings, spikes, modules, gates, repeated structures.
That is huge.
2. His “systems thinking” is now visible
The drawing gives us visual evidence of something we’ve suspected from the block structures and music: Gavin organizes thought through nested worlds.
He doesn’t just make a thing.
He makes a thing with:
a center
rules of access
protective edges
repeating parts
directional pathways
tension zones
structural symmetry
That becomes a category in the portfolio: visual-symbolic systems construction.
3. Future drawings should be treated like compositions
I would now analyze his drawings the way we analyze his songs:
What is the emotional center?
Where is the tension?
What repeats?
What protects the core?
What changes from one side to the other?
Where is the “pedal” equivalent — the part that creates resonance or atmosphere?
Is the system open, closed, defended, or expanding?
This one is almost a “score,” just in spatial form.
4. It strengthens the Project Zero / constructionism angle
This picture supports the idea that Gavin is externalizing thinking through artifacts. That’s very Papert/Project Zero: learning made visible, not by answering questions, but by constructing something that carries thought.
The key phrase going forward might be:
Gavin does not merely express ideas.
He constructs symbolic environments where ideas can live.
5. It gives us a new portfolio section
I’d start a section called something like:
Gavin’s Visual Worlds: Spatial-Symbolic System Maps
This drawing can be Entry 001 in that section.
Working title:
“The Protected Core”
or
“Central System With Boundary Fields”
My honest read: this is one of the clearest non-musical artifacts we’ve seen showing that the same child who composes emotional worlds at the piano also draws contained symbolic architectures on paper.
The map is migrating.

What do you think of GTD, i.e., externalizing the state in writing?
vibe coding just exposed a fatal flaw in how most indie makers build products.
most solopreneurs compress their build process into shipping. you get the "what" but not the "why." invisible reasoning behind every product decision gets lost.
the fix? treat every feature like a first-principles question.
- why does this user have this problem
- what is the cheapest way to validate it
- what breaks if you remove this
the result: a decision chain where every build choice traces back to a real user need.
here's where it shifts everything - instead of asking "what should i build next," you retrieve ALL the reasoning that justifies it: from pain point to market signal to retention mechanic to monetization path.
no more shipping on vibes alone. every feature is verifiable.
every decision has a chain.
this isn't just better building. it's externalizing the invisible logic that separates products that grow from products that die.
stateless is a dream until you have to rebuild the entire context graph on every request. the complexity doesn't vanish, it just turns into a massive latency tax and a higher api bill. externalizing state to something like redis or a persistent session store is the only way to keep agents fast without burning through credits re-sending the same history.
Architecture isn't about removing state. It's about deciding who carries the weight
Pure stateless pushes complexity to the client
Pure stateful pushes complexity to the server
Neither scales cleanly past a certain point
The real move is externalizing state, Redis, edge caches, session stores
Same problem, better placement
#SoftwareEngineering
Spring profiles was that moment for me too. The next unlock is externalizing secrets into env vars or a vault instead of keeping them in application-prod.yml, that is when deploys start feeling actually safe.
The main limitation of current agents becomes clear once tasks extend over longer periods of time
It’s not that they fail at individual steps, but that they struggle to stay consistent across an evolving workflow
As projects move through setup, implementation, experimentation, and debugging, alignment starts to break down
Context drifts, intermediate decisions aren’t preserved clearly, and the overall process becomes harder to track or trust
Framing this as a problem of orchestration and state continuity helps clarify what’s actually missing
It’s less about generating outputs in isolation and more about coordinating multiple stages while maintaining a clean and persistent record of what’s happening
The file based coordination approach highlights how important explicit structure is
By externalizing state, it reduces ambiguity and keeps each step grounded in the same context
The performance drop without it makes it clear this kind of structure isn’t optional, it’s foundational
This is where the direction of @ritualnet Network starts to feel closely aligned
As AI systems take on longer and more complex workflows, trust shifts from single outputs to the entire execution process
Being able to verify how results are produced, what steps were taken, and how state was maintained becomes increasingly important
It’s not just about running agents, but about making their behavior transparent and verifiable over time
And that kind of infrastructure becomes more critical as these systems move closer to real world use cases
[Weekly Ritual Digest 4 | Automating the Long Game: Autonomous ML Research Agents]
@ritualnet @ritualfnd @ritualnet_korea
Hello!
The final topic of this week's digest introduces a study that analyzes why agents break down during ML research engineering spanning hours or days, and proposes a new system to overcome this.
1. Where agents break: Long-horizon workflows
Long-horizon ML research engineering—spanning specification, setup, implementation, experiments, and debugging—is exactly where current AI agents break down. On a 48-hour budget, ML PhDs hit a 41% success rate on PaperBench, while the top agent achieved only 21%.
2. Orchestration and state-continuity
To solve this, the authors proposed an AI scientist framework built on the principle of treating long-horizon performance as a dual problem of orchestration and state-continuity. Agents must coordinate across heterogeneous stages while preserving the evolving project state with absolute clarity.
3. The immense value of File-as-Bus coordination
This approach improved PaperBench scores by 10.54 points over baselines and reached 81.82% 'Any Medal' on MLE-Bench Lite. In one 23-hour run, it successfully executed 74 experiment cycles. Crucially, removing the 'File-as-Bus' design caused PaperBench scores to drop by 6.41 points and MLE-Bench by 31.82 points, proving its fundamental value.
4. Wrapping up this week's digest
This week, we explored the nuanced differences in pretraining optimization, the handling of conflicting instructions, the errors of exploration vs. exploitation, and the automation of long-horizon research projects.
As AI takes on increasingly complex and autonomous roles, the need for a decentralized infrastructure to transparently verify and trust these intricate execution processes becomes critical. This is exactly why the secure AI ecosystem being built by Ritual Network is so essential. We will return next week with more insightful research. Have a great weekend!
---
[Weekly Ritual Digest 4 | 긴 호흡의 작업 자동화: 자율적인 ML 리서치 에이전트]
안녕하세요?
이번 주 다이제스트의 마지막 주제는 수 시간에서 수일이 걸리는 머신러닝 리서치 엔지니어링 과정에서 에이전트가 무너지는 원인을 분석하고 이를 극복하는 새로운 시스템을 제안한 연구입니다.
1. 긴 호흡의 작업에서 무너지는 AI
설계부터 환경 설정, 구현, 실험, 디버깅에 이르는 장기(Long-horizon) 프로젝트는 현재 AI 에이전트들의 가장 큰 약점입니다. 48시간의 예산이 주어진 실제 PaperBench 평가에서 머신러닝 박사 과정 학생들은 41%의 성공률을 보였지만, 최고 성능의 에이전트는 21%에 그쳤습니다.
2. 오케스트레이션과 상태 연속성
연구진은 이 문제를 해결하기 위해 AI 과학자 시스템을 제안했습니다. 긴 호흡의 작업을 다양한 역할을 조율하는 오케스트레이션과 진화하는 프로젝트의 상태를 명확하게 보존하는 연속성의 문제로 접근한 것입니다.
3. 파일 기반(File-as-Bus) 조율의 가치
이 시스템은 베이스라인 대비 PaperBench에서 10.54점을 향상시켰고, MLE-Bench Lite에서는 81.82%의 메달 획득률(Any Medal)을 달성했습니다. 한 번의 23시간 실행 동안 74번의 실험 주기를 수행하기도 했습니다. 특히 이 구조의 핵심인 파일 기반 조율(File-as-Bus)을 제거하자 성능이 대폭 하락하며 그 중요성이 명확히 입증되었습니다.
4. 이번 주 다이제스트를 마치며
이번 주에는 사전 학습의 디테일한 최적화, 다중 지시어의 충돌, 탐색과 활용의 오류, 그리고 장기 리서치 프로젝트의 자동화까지 에이전트가 현실의 복잡성에 대응하기 위해 진화해 나가는 과정들을 살펴보았습니다.
이처럼 AI의 역할과 자율성이 고도화될수록, 그 복잡한 실행 과정을 투명하게 검증하고 신뢰할 수 있는 탈중앙화된 인프라의 필요성 역시 커지고 있습니다. 리추얼 네트워크가 구축하는 안전한 AI 생태계가 더욱 기대되는 이유입니다. 오늘 글은 여기서 줄이겠습니다. 그럼 다들 행복한 주말 보내세요!
@joshsimenhoff @mongdiny7 @Jez_Cryptoz @dunken9718
![mechoorial's tweet photo. [Weekly Ritual Digest 4 | Automating the Long Game: Autonomous ML Research Agents]
@ritualnet @ritualfnd @ritualnet_korea
Hello!
The final topic of this week's digest introduces a study that analyzes why agents break down during ML research engineering spanning hours or days, and proposes a new system to overcome this.
1. Where agents break: Long-horizon workflows
Long-horizon ML research engineering—spanning specification, setup, implementation, experiments, and debugging—is exactly where current AI agents break down. On a 48-hour budget, ML PhDs hit a 41% success rate on PaperBench, while the top agent achieved only 21%.
2. Orchestration and state-continuity
To solve this, the authors proposed an AI scientist framework built on the principle of treating long-horizon performance as a dual problem of orchestration and state-continuity. Agents must coordinate across heterogeneous stages while preserving the evolving project state with absolute clarity.
3. The immense value of File-as-Bus coordination
This approach improved PaperBench scores by 10.54 points over baselines and reached 81.82% 'Any Medal' on MLE-Bench Lite. In one 23-hour run, it successfully executed 74 experiment cycles. Crucially, removing the 'File-as-Bus' design caused PaperBench scores to drop by 6.41 points and MLE-Bench by 31.82 points, proving its fundamental value.
4. Wrapping up this week's digest
This week, we explored the nuanced differences in pretraining optimization, the handling of conflicting instructions, the errors of exploration vs. exploitation, and the automation of long-horizon research projects.
As AI takes on increasingly complex and autonomous roles, the need for a decentralized infrastructure to transparently verify and trust these intricate execution processes becomes critical. This is exactly why the secure AI ecosystem being built by Ritual Network is so essential. We will return next week with more insightful research. Have a great weekend!
---
[Weekly Ritual Digest 4 | 긴 호흡의 작업 자동화: 자율적인 ML 리서치 에이전트]
안녕하세요?
이번 주 다이제스트의 마지막 주제는 수 시간에서 수일이 걸리는 머신러닝 리서치 엔지니어링 과정에서 에이전트가 무너지는 원인을 분석하고 이를 극복하는 새로운 시스템을 제안한 연구입니다.
1. 긴 호흡의 작업에서 무너지는 AI
설계부터 환경 설정, 구현, 실험, 디버깅에 이르는 장기(Long-horizon) 프로젝트는 현재 AI 에이전트들의 가장 큰 약점입니다. 48시간의 예산이 주어진 실제 PaperBench 평가에서 머신러닝 박사 과정 학생들은 41%의 성공률을 보였지만, 최고 성능의 에이전트는 21%에 그쳤습니다.
2. 오케스트레이션과 상태 연속성
연구진은 이 문제를 해결하기 위해 AI 과학자 시스템을 제안했습니다. 긴 호흡의 작업을 다양한 역할을 조율하는 오케스트레이션과 진화하는 프로젝트의 상태를 명확하게 보존하는 연속성의 문제로 접근한 것입니다.
3. 파일 기반(File-as-Bus) 조율의 가치
이 시스템은 베이스라인 대비 PaperBench에서 10.54점을 향상시켰고, MLE-Bench Lite에서는 81.82%의 메달 획득률(Any Medal)을 달성했습니다. 한 번의 23시간 실행 동안 74번의 실험 주기를 수행하기도 했습니다. 특히 이 구조의 핵심인 파일 기반 조율(File-as-Bus)을 제거하자 성능이 대폭 하락하며 그 중요성이 명확히 입증되었습니다.
4. 이번 주 다이제스트를 마치며
이번 주에는 사전 학습의 디테일한 최적화, 다중 지시어의 충돌, 탐색과 활용의 오류, 그리고 장기 리서치 프로젝트의 자동화까지 에이전트가 현실의 복잡성에 대응하기 위해 진화해 나가는 과정들을 살펴보았습니다.
이처럼 AI의 역할과 자율성이 고도화될수록, 그 복잡한 실행 과정을 투명하게 검증하고 신뢰할 수 있는 탈중앙화된 인프라의 필요성 역시 커지고 있습니다. 리추얼 네트워크가 구축하는 안전한 AI 생태계가 더욱 기대되는 이유입니다. 오늘 글은 여기서 줄이겠습니다. 그럼 다들 행복한 주말 보내세요!
@joshsimenhoff @mongdiny7 @Jez_Cryptoz @dunken9718](https://pbs.twimg.com/media/HGosjyeagAAHm8o.png)
Ritual externalizing compute across chains, dependency layer beats competition that's the real infra play
If you’ve ever run a prover cluster, you know , the hard part isn’t proving. It’s everything around it.
That’s exactly the layer @fermah_xyz is trying to abstract away.
We tend to frame ZK as a cryptography problem.
In practice, it’s an infrastructure and coordination problem:
orchestration ,resource allocation ,system design around proving
The math is solid , The systems are where things break.
Right now , most proving stacks look like this:
tightly coupled prover setups
infra heavy deployments
underutilized hardware
fragile pipelines across components
You’re not just generating proofs , you’re running a system.
And the real constraint isn’t raw compute.
It’s coordinating across:
heterogeneous hardware (CPU / GPU / FPGA)
multiple proving systems
dynamic workloads
There’s no shared layer handling this cleanly.
So every team ends up rebuilding the same thing:
provisioning infra
writing custom schedulers
optimizing locally
Which leads to fragmentation , exactly where composability should exist.
A more scalable model is to treat proving as a routable workflow.
Not:
Which prover runs this job?
But:
What computation needs to be proven—and who can execute it best?
That shift separates concerns:
Seekers define workloads
Prover Nodes supply compute
a Match maker coordinates execution
Now proving becomes something you can route, not hardwire.
The Match maker layer does more than scheduling.
It handles:
workload decomposition
hardware aware routing
latency vs cost tradeoffs
multi stage proving flows
All while maximizing utilization across the network.
This is what universal means in practice:
any proof system
any chain
any VM
any workload shape
Because fragmentation at the proving layer limits everything upstream.
Take rollups as an example.
Externalizing proving isn’t just about cost , it’s architectural:
fewer long lived actors , less cross component messaging , clearer execution boundares
From actor graphs → to dataflow pipelines.
That simplifies the system:
proving becomes function like
failure domains are easier to reason about
infra overhead drops
You stop managing prover fleets and start invoking proofs as a primitive.
There’s also a second order effect:
Aggregated demand → higher utilization → lower marginal cost
Which expands what’s economically viable to prove.
For different actors:
Builders → focus on what to prove
Provers → compete on performance & reliability
Researchers → test systems under real load
@fermah_xyz @flashcaster @0xTribal @vanishree_rao
The most productive thing I do every week isn't code, marketing, or meetings.
it's spending 10 minutes speaking into a mic with no agenda.
no structure. no bullet points. just talking.
the clarity that comes from externalizing your chaos is underrated.
Last Seen Hashtags on Sotwe
Trends for you
Most Popular Users

Elon Musk 
@elonmusk
240.3M followers

Barack Obama 
@barackobama
118.6M followers

Cristiano Ronaldo 
@cristiano
114.4M followers

Donald J. Trump 
@realdonaldtrump
111.8M followers

Narendra Modi 
@narendramodi
106.9M followers

Rihanna 
@rihanna
98.7M followers

NASA 
@nasa
92.2M followers

Justin Bieber 
@justinbieber
91.7M followers

KATY PERRY 
@katyperry
89.9M followers

Taylor Swift 
@taylorswift13
83.9M followers

Lady Gaga 
@ladygaga
75.4M followers

Virat Kohli 
@imvkohli
73.3M followers

Kim Kardashian 
@kimkardashian
70.8M followers

YouTube 
@youtube
68.6M followers

Neymar Jr 
@neymarjr
66.3M followers

Bill Gates 
@billgates
65.2M followers

Selena Gomez 
@selenagomez
63M followers

The Ellen Show
@theellenshow
62.1M followers

CNN 
@cnn
61.7M followers

X 
@x
60.6M followers


















