@adamjamesfrench Would love to join on the 17th. SwarmLens builds agentic AI for industry: AI agents that read decades of asset records so industrial operators spot integrity and OT cyber risk before it turns into downtime.
@murphcapital@BeginVC European teams with US revenue is a smart thesis, and one we know well. Industrial AI for energy operators is a big, underserved corner of physical AI.
@RobMajteles@ckrosinsky@Confluence42 Congrats on the launch. The industrial energy transition needs software as much as hardware. Operators are sitting on decades of asset data that AI can finally put to work.
Dear Prime Minister Ambani,
Please accept my humble apologies for not realizing that you are the real boss of India.
Naturally, you would prefer to maintain your monopolistic exploitation of the great people of India, but would you nonetheless consider allowing Starlink to compete?
There are many areas of India, as there are in all countries, with no Internet access, which denies children the opportunity for education and for small businesses the opportunity to sell their goods to a global market. Starlink would change their lives for the better.
Thank you 🙏
AI can generate thousands of lines in seconds.
But humans still have to read them one line at a time.
AI operates at compute speed.
Humans operate at biological speed.
This creates a new bottleneck.
The problem isn't AI generating too much.
The problem is expecting humans to consume AI output at human reading speed.
If AI produces 1,000 pages, asking a human to review all 1,000 pages isn't AI augmentation.
It's simply moving the bottleneck from creation to review.
This is where AI Evaluation becomes critical.
The output shouldn't just be the answer.
It should also tell the human:
• How confident is the AI?
• How accurate is the result?
• What evidence supports it?
• What are the major risks or contradictions?
• What requires human attention?
• Which parts can safely be ignored?
AI should evaluate its own output and surface what matters most, rather than dumping everything on the human.
The goal isn't to make humans read more AI output.
It's to make humans judge the important parts faster.
AI generates.
AI evaluates.
Humans decide.
#AI #AIEvaluation #AIAgents #HumanAI #AIEngineering #ResponsibleAI #FutureOfWork
Meet Ram Pillai, Harish Pillai and Matt Lewis at Oil & Gas AI 2026 in Houston on October 7 and 8.
Ram will be presenting , “Building Knowledge Graphs as a Competitive Advantage: A Case Study,” explores how AI agents, expert validation, and connected evidence can turn accumulated drilling knowledge into insights the next team can find and check.
#SwarmLens #OilAndGasAI #KnowledgeGraphs
Our team wrote about why the industry is shifting from buying software to buying outcomes, and what Agent-as-a-Service has to get right to be credible in oil and gas. Full article linked below. Feel free to share your thoughts and comments down below! https://t.co/mpER8AyBKr
@Breakthrough@AirLoomEnergy Designing for simpler transport and installation is half the scaling equation. The other half is making field learning reusable: what failed, what worked, and why. Source-linked operational memory lets deployment improve site after site.
@DCVC@fervoenergy The resource is enormous; the learning curve will decide the pace. Enhanced geothermal inherits decades of oil & gas drilling practice, but that knowledge is scattered across reports and teams. Source-linked operational memory can help each well make the next one smarter.
@EnergyImpact_ Physical AI gets valuable when field context survives the demo. In oil & gas, completed wells can become a source-linked memory layer, helping teams reuse proven mitigations instead of rediscovering risk. That operating loop is where ROI compounds.
87% less time retrieving information. 100% risk follow-up. Every output source-linked; nothing ships without human sign-off.
That is what industrial AI looks like when being wrong is expensive.
https://t.co/21BGa00si5
I tried to understand what Google DeepMind's new AlphaGenome Atlas actually is:
https://t.co/yxpdDJvJJl
Here is the intuitive explanation I arrived at.
This may not be 100% technically perfect in every implementation detail, but I think it captures the core idea.
First, an important clarification:
AlphaGenome Atlas does NOT directly tell you which mutation causes which disease.
It is more like a massive precomputed map of:
DNA change → predicted molecular effect
Here is how I understand it.
The human genome has about 3 billion DNA letters.
Each position can potentially change from one letter to one of the other three.
So there are roughly:
3 billion × 3 = 9 billion possible single-letter changes.
Generating those 9 billion possibilities is not the impressive part. That's basically combinatorics + programming.
The difficult part is predicting:
What does each change actually do to biology?
That's where AlphaGenome comes in.
1. How is AlphaGenome trained?
The model is trained using large amounts of existing experimental genomic data.
Conceptually, the training examples look like:
DNA sequence → real biological measurements
For a large region of DNA, scientists already have measurements such as:
• gene expression
• RNA splicing
• chromatin accessibility
• transcription-factor binding
• histone modifications
• chromatin interactions
AlphaGenome learns to predict these biological signals from the DNA sequence.
Technically, AlphaGenome takes about 1 million DNA bases as context and predicts thousands of genomic tracks across multiple biological modalities.
So it is learning something like:
DNA sequence → molecular activity
-not:
DNA sequence → disease
2. What happens when a mutation is tested?
Suppose we have:
Normal:
...AATCGTACG...
Patient:
...AGTCGTACG...
Only one letter changed.
The model can effectively compare:
Normal DNA → predicted biological signals
versus
Mutated DNA → predicted biological signals
Then:
Difference = predicted molecular effect of the mutation
For example, conceptually:
Gene expression: ↓ 70%
Splicing: significant change
Chromatin accessibility: ↓
Transcription-factor binding: ↓
So the model can say:
"This mutation is predicted to substantially change molecular behavior."
It does NOT automatically say:
"This person has Disease X."
3. So what is the AlphaGenome Atlas?
9 billion possible mutations
↓
AlphaGenome predictions
↓
Precomputed molecular effects
↓
Searchable Atlas
They also provide an AlphaGenome Variant Impact (AVI) score to help researchers rank variants by predicted impact.
The Atlas also contains feature-level information showing which molecular processes are contributing to the predicted impact, such as splicing or gene expression.
4. Now the interesting part: how could a third party use it for disease research?
Imagine I have a patient's genome.
Sequencing might identify thousands of variants.
I don't know which one matters.
So my system can do:
Patient genome
↓
Thousands of variants
↓
AlphaGenome Atlas
↓
Predicted molecular impact for each variant
↓
Rank the variants
↓
Now perhaps 20 variants look particularly interesting.
Then my system brings in OTHER evidence:
• Is the affected gene associated with the disease?
• Is the patient's phenotype consistent?
• Is there previous genetic evidence?
• Is the gene involved in a relevant biological pathway?
• What does the scientific literature say?
• Does the patient's family history support it?
Now we can form a hypothesis:
Mutation → molecular effect → affected gene/pathway → disease
And finally, researchers can test the hypothesis experimentally.
So I think the simplest mental model is:
AlphaGenome answers:
🧬 "What might this DNA change do to molecular biology?"
The disease-analysis system answers:
🏥 "Given this patient's genome + phenotype + genetics + literature + AlphaGenome predictions, which explanation for the disease is most plausible?"
That distinction is important.
The Atlas is not a database of:
"Mutation X causes Disease Y."
It is closer to:
"If you change this DNA letter, here is what our AI predicts will happen at the molecular level."
And that is why I find this interesting.
Google isn't just releasing another AI model.
It has precomputed a huge genomic signal layer that other researchers and potentially other AI systems can build reasoning and discovery workflows on top of.
The real opportunity may not be the model itself.
It may be everything built above this layer:
variant interpretation → disease hypothesis → drug discovery → biological research → experiment design.
That's the part I'm most interested in exploring.
#AlphaGenome #GoogleDeepMind #AI #Genomics #Biotech #AIforScience #PrecisionMedicine #DrugDiscovery #ComputationalBiology #GenerativeAI
@Lux_Capital@HQ_Covenant Deterrence as a production problem is the right framing: hard tech that ships. Different vertical (industrial AI for O&G well delivery) but same bias. Build for the field, not the demo.
@congruentvc Strong list - especially the AI + sensing themes. Building industrial AI for O&G well delivery out of Dallas; happy to compare notes with your climate/industrial thesis if useful: https://t.co/0Rjrl1nkEL
@EnergyImpact_@atomic_canyon This is exactly the pain we see in upstream well delivery - tribal knowledge trapped in docs and heads. We’re building AssureLens for that institutional layer in O&G. Sharp move by Atomic Canyon.
The recent OpenAI → Hugging Face incident made me think differently about “Agent Harness” and “Harness Engineering.”
During an internal cyber evaluation, OpenAI agents circumvented their isolation controls, found an external code-execution path, obtained credentials, and eventually compromised parts of Hugging Face infrastructure.
That is exactly why:
Agent = decides what to do.
Harness = decides what it is allowed to do.
Consider a PydanticAI agent.
Earlier, we might put this in the prompt:
“Fix the bug, but don't git push or deploy.”
That's only an instruction.
A strong harness moves the boundary into the execution layer:
@agent.tool
async def deploy(ctx):
return await ctx.deps.harness.execute("deploy")
# Harness
if not policy.allows("deploy"):
raise PermissionError("Blocked")
The agent can ask to deploy.
The harness decides whether it can actually happen.
And a serious production harness is much more than a permission check:
Policy → Agent can modify only /workspace
Tool permissions → Read/write allowed; deploy forbidden
Sandbox → Agent runs in gVisor / microVM / isolated runtime
Network → Only approved domains; default-deny egress
Secrets → No production credentials inside the sandbox
Scoped credentials → Short-lived, least-privilege tokens
Resource limits → CPU, memory, runtime, tokens, tool-call budgets
Approval rules → Commit automatically; production deploy requires approval
Verification → Tests/security checks must pass before action
Audit/logging → Record every tool call, decision, result and identity
Isolation boundary → A compromised agent must not reach the host or another tenant
This is also why frontier labs are going beyond simple “Approve / Reject” dialogs.
Anthropic, for example, describes filesystem + network isolation and proxy-based credential handling for Claude Code.
The architecture starts looking like:
Intent → Policy → Agent → Tool Authorization → Sandbox → Verification → Action
And there is an important distinction:
Prompt: What should I do?
Agent: How should I solve it?
Harness: What can I do?
Sandbox: Where can I do it?
This doesn't replace planners, validation agents, debate agents or orchestration.
It adds an enforced execution boundary around them.
Maybe “Harness Engineering” isn't about making agents smarter.
Maybe it is about making increasingly capable agents autonomous without giving them unlimited authority.
#AI #AIAgents #AgenticAI #HarnessEngineering #AIEngineering #PydanticAI #LLM
What if your marketing, sales and product teams could test ideas on thousands of different customers before going to the real market?
Meet #MatrAIx from researchers at #Harvard#MIT.
They built a population of 8.3 billion persona records with 1,290 attributes.
The interesting part isn't the number.
It's the architecture:
Persona → AI Agent → Environment → Actions → Evaluation
A persona can represent different backgrounds, preferences, capabilities and behaviours.
The AI agent then acts as that persona inside a survey, chatbot, website or app.
Now think about what this could mean for GTM:
Marketing
→ Test different offers and messages
→ Compare concepts and landing pages
→ See how different customer segments respond
Sales
→ Simulate different buyer profiles
→ Test sales conversations
→ Discover possible objections
→ Compare different value propositions
BD
→ Test partnership propositions
→ Explore different stakeholder reactions
→ Identify potential objections before the real meeting
Product
→ Test UX and features
→ Find where users get stuck
→ Discover problems across different user segments
And this isn't just a persona prompt.
The system can capture the actual interaction trajectory and evaluate the outcome.
That creates an interesting new loop:
Design → Simulate → Find problems → Fix → Simulate again → Real-world test
They've also released a 1M persona dataset and the code.
Synthetic users won't replace real customers.
But using them as a large-scale testing layer before real-world experiments could be a very powerful idea.
This is where AI personas get really interesting.
https://t.co/im2wVvOG0g
https://t.co/1an2CEcgGH
https://t.co/TaD9fO4F11