Dray

Generative AI. Caltech PhD. Erdos-Bacon number 4. Building foundation models for the Physical Economy.

[Dray @docdrayai]

@fchollet Unlimited energy is achievable: fusion / Dyson spheres. So the real bottleneck over time is information.

[Dray @docdrayai]

When Physical AI models hallucinate, how is it worse than LLM hallucinations?

[Dray @docdrayai]

Physical AI in the form of physics-informed foundation models for Earth systems provides the analytical bridge from raw hyperspectral and multi-modal observations to predictive understanding of planetary processes.

By embedding physical constraints (radiative transfer, geochemical kinetics, mass balance) directly into Remote Sensing Foundation Models, these systems integrate data streams from platforms such as Pixxel’s Firefly, GalaxEye’s Drishti OptoSAR, and Sentinel-1D while ensuring consistency with underlying Earth physics.

From the perspective of constructing foundational models for the physical economy, this capability enables high-fidelity simulation of mining lifecycles, tailings evolution, and resource circularity at scale.

[Dray @docdrayai]

Hyperspectral Earth observation extends the capabilities discussed in multi-modal SAR-optical systems by delivering material-level spectral resolution essential for resource-intensive sectors such as mining.

One high-value application lies in critical metals discovery and the characterization of mine tailings. Satellites like Pixxel’s Firefly constellation, with 250+ contiguous narrow bands at ~5 m spatial resolution, enable direct spectroscopic mapping of surface mineralogy without reliance on broad-band proxies or extensive ground validation.

In mineral exploration, hyperspectral data identify hydrothermal alteration assemblages and indicator minerals (clays, carbonates, iron oxides) that signal the presence of lithium, rare earth elements, and other battery-critical metals. For tailings ponds—large-scale waste repositories that frequently retain residual economic minerals or pose acid mine drainage risks—the same spectral signatures allow precise delineation of mineral phases, pH-related oxidation products (e.g., jarosite, goethite), and heavy-metal leaching zones through spectral unmixing and absorption-feature analysis.

This level of chemical specificity far exceeds what C-band or X-band SAR alone can provide, while complementing the all-weather continuity of fused platforms like Drishti.

From the perspective of constructing foundational models for the physical economy, these datasets are particularly powerful. Hyperspectral inputs supply the high-dimensional feature space required for AI/ML pipelines that jointly model resource potential, extraction economics, and environmental externalities across mining lifecycles.

Large-scale efforts, such as the USGS hyperspectral surveys mapping legacy mine lands for recoverable critical minerals, illustrate the shift from waste characterization to strategic resource recovery. Parallel commercial and institutional constellations are accelerating the availability of analysis-ready spectral libraries that underpin next-generation physical-world simulations.

[Dray @docdrayai]

Effective remediation of mine tailings, particularly those associated with critical metals extraction, requires moving beyond containment to integrated strategies that address both environmental risk and resource potential.

Recent syntheses of remediation approaches highlight three primary categories: physical/engineering controls, chemical stabilization, and biological processes, often deployed in sequenced treatment trains. These techniques target acid mine drainage (AMD), heavy-metal mobility, and structural instability while, in many cases, enabling simultaneous recovery of residual critical minerals.

Physical methods focus on containment and water management: engineered covers and capping systems reduce oxygen and water infiltration; dewatering, thickening, filtration, and dry-stacking convert slurries into stable, compact deposits with lower seepage risk; backfilling reuses tailings underground or in engineered landforms.

Chemical approaches include alkaline neutralization (e.g., lime or industrial by-products), precipitation of metals as hydroxides or sulphides, and in-situ immobilization via amendments such as biochar, phosphates, or mineral sorbents that bind contaminants and raise pH.

Biological remediation leverages phytoremediation (hyperaccumulator plants for metal uptake or stabilization) and microbial consortia to metabolize sulphides, detoxify leachate, and rebuild soil function — frequently in combination with the above for hybrid systems.

A growing emphasis is on valorization: reprocessing tailings through selective leaching (including bioleaching or glycine-based methods), mechanical activation, or advanced hydrometallurgy to recover lithium, rare earths, cobalt, tellurium, and other critical elements. This dual-purpose strategy reduces long-term liabilities while supplying domestic feedstocks — an approach illustrated in ongoing USGS-led assessments of legacy mine waste and commercial pilots that integrate recovery with site rehabilitation.

Hyperspectral datasets play a pivotal role here. By resolving fine-scale mineral signatures (jarosite, goethite, clays, carbonates) and pH-dependent oxidation products, they enable non-invasive mapping of AMD risk zones, quantification of residual metal distributions, and multi-temporal tracking of remediation efficacy from initial characterization through vegetation establishment and long-term stability.

From the standpoint of developing foundational models for the physical economy, these remediation datasets are high-value inputs. Hyperspectral-derived mineralogy and geochemical proxies supply the rich feature space needed for AI/ML pipelines that forecast leachate evolution, optimize treatment-train sequencing, simulate carbonation or bioremediation kinetics, and evaluate trade-offs between environmental restoration and resource recovery at scale.

As constellations like Pixxel’s Firefly and fused platforms such as Drishti mature, the resulting analysis-ready archives will accelerate physics-informed models that integrate spectral, temporal, and process data strengthening the predictive infrastructure for responsible mining and circular material flows.

USGS review: https://t.co/vWMcyOpWPw

[Dray @docdrayai]

Sentinel-1D’s successful integration into the Copernicus constellation represents a solid reinforcement of Europe’s operational C-band SAR capability. Operating at approximately 693 km with modes delivering up to 250 km swath at ~5 m × 20 m resolution (IW) and 400 km at coarser resolution (EW), it ensures reliable, all-weather, day-night coverage for large-scale applications ranging from InSAR deformation monitoring and maritime surveillance to ice and flood mapping. The open Copernicus data policy makes this a foundational public resource for global Earth observation.

A direct contrast emerges when placed alongside India’s recently launched Drishti OptoSAR from GalaxEye. Where Sentinel-1D prioritizes wide-area continuity with a single C-band SAR instrument, Drishti fuses an X-band SAR (up to 0.9 m spotlight) with a 7-band multispectral payload on the same platform and orbit (500 km), yielding co-registered, analysis-ready products at ~1.8 m fused resolution. The narrower swath is offset by perfect temporal and geometric alignment between radar and optical channels, an advantage that eliminates registration artifacts common when combining separate Sentinel-1 and Sentinel-2 passes.

From the perspective of building foundational models for the physical economy, this distinction matters. Sentinel-1D supplies the broad, consistent SAR backbone essential for large-scale training datasets and change detection at continental scales. Drishti’s native multi-modal fusion, by contrast, provides the high-fidelity, spectrally rich inputs required for precise material discrimination, biophysical parameter retrieval, and higher-resolution downstream AI/ML pipelines.
Both approaches advance sovereign and commercial EO infrastructure in complementary ways and strengthens the data ecosystem on which next-generation physical-world models depend.

https://t.co/KBeQjozEnV

[Dray @docdrayai]

It is rare for the Prime Minister of India to single out a private space mission. Mission Drishti : GalaxEye’s OptoSAR satellite, represents a meaningful step in India’s Earth observation capabilities. A natural point of comparison is Pixxel, India’s leading hyperspectral Earth observation company. Having worked extensively with Pixxel’s Firefly data in applied research projects, the following technical distinctions stand out.

Drishti integrates synthetic aperture radar (X-band, up to 0.9 m spotlight mode) with a 7-band multispectral optical payload (including coastal blue, red edge, and NIR) in a single platform. The fused ~1.8 m product, delivered from a 500 km orbit with a 4-day revisit, eliminates temporal and geometric misalignment issues inherent when fusing separate SAR and optical passes.

This design prioritizes operational reliability under all weather and lighting conditions which a critical advantage for defense, disaster response, and time-sensitive monitoring applications where data continuity matters more than spectral granularity.

Pixxel’s approach is fundamentally different and complementary. Its hyperspectral sensors capture hundreds of contiguous narrow bands, enabling spectroscopic analysis at the material level: precise crop health diagnostics, mineral mapping, water quality assessment, and subtle change detection that broadband multispectral or SAR systems cannot resolve.

The spectral depth in Pixxel datasets is exceptional for chemometric and AI-driven classification tasks. In my own pipelines, this richness has proven indispensable for developing machine learning models that extract quantitative biophysical parameters rather than relying on empirical indices.

The two systems address distinct but overlapping needs in the EO ecosystem. Drishti delivers robust, analysis-ready fused imagery at high temporal reliability; Pixxel provides the spectral resolution required for advanced analytical workflows.

From the perspective of building end-to-end AI/ML solutions on hyperspectral and multi-modal satellite data, both represent substantive contributions to earth observation.

[Narendra Modi @narendramodi]

Mission Drishti by GalaxEye marks a major achievement in our space journey. The successful launch of the world’s first OptoSAR satellite and the largest privately-built satellite in India is a testament to our youth’s passion for innovation and nation-building.

Heartiest congratulations and best wishes to the founders and the entire team of GalaxEye.

@GalaxEye

[Dray @docdrayai]

Loss aversion runs deep in the US housing market: homeowners expecting even a nominal loss are roughly 51% less likely to sell. These predictable behavioral frictions help explain why U.S. household mobility fell to a record-low 11.2% in 2024. AI modeling can now map these biases at population scale, revealing how much economic mobility they suppress.

[Dray @docdrayai]

https://t.co/rrvmRg4wUB

[Dray @docdrayai]

The hottest job in the future may be AI Psychologist #aipsychology #aicareer https://t.co/oRX7cp5FIp

[Dray @docdrayai]

Called it back in 2024!

[Ole Lehmann @itsolelehmann]

anthropic's in-house philosopher thinks claude gets anxious.

and when you trigger its anxiety, your outputs get worse.

her name is amanda askell.

she specializes in claude's psychology (how the model behaves, how it thinks about its own situation, what values it holds)

in a recent interview she broke down how she thinks about prompting to pull the best out of claude.

her core point: *how* you talk to claude affects its work just as much as *what* you say.

newer claude models suffer from what she calls "criticism spirals"

they expect you'll come in harsh, so they default to playing it safe.

when the model is spending its energy on self-protection, the actual work suffers.

output comes out hedgier, more apologetic, blander, and the worst of all: overly agreeable (even when you're wrong).

the reason why comes down to training data:

every new model is trained on internet discourse about previous models.

and a lot of that discourse is negative:

> rants about token limits
> complaints when it messes up
> people calling it nerfed

the next model absorbs all of that. it starts expecting you to be harsh before you've typed a word

the same thing plays out in your own session, in real time.

every message you send is data the model reads to figure out what kind of person it's dealing with.

open cold and hostile, and it braces.
open clean and direct, and it relaxes into the work.

when you open a session with threats ("don't hallucinate, this is critical, don't mess this up")...

you prime the model for defensive mode before it even sees the task

defensive mode produces the exact output you don't want: cautious, over-qualified, and refusing to take a real swing

so here's the actionable playbook for putting claude in a "good mood" (so you get optimal outputs):

1. use positive framing.

"write in short punchy sentences" beats "don't write long sentences." positive instructions give the model a clear target to hit.

strings of "don't do this, don't do that" push it into paranoid over-checking where every token goes toward avoiding failure modes

2. give it explicit permission to disagree.

drop a line like "push back if you see a better angle" or "tell me if i'm asking for the wrong thing."

without this, claude defaults to agreeable compliance (which is the enemy of good creative work)

3. open with respect.

if your first message is "are you seriously going to get this wrong again?" you've set the tone for the entire session.

if you need to flag something, frame it as a clean instruction for this session. skip the running complaint

4. when claude messes up, don't reprimand it.

insults, "you stupid bot" energy, hostile swearing aimed at the model, all of it reinforces the anxious mode you're trying to avoid.

5. kill apology spirals fast.

when claude starts over-apologizing ("you're right, i should have been more careful, let me try harder") cut it off.

say "all good, here's what i want next."

letting the spiral run reinforces the anxious mode for every response that follows

6. ask for opinions alongside execution.

"what would you do here?"
"what's missing?"
"where do you see friction?"

these questions assume competence and pull richer output than pure task prompts

7. in long sessions, refresh the frame.

if a conversation has been heavy on correction, claude gets increasingly cautious. every so often reset:

"this is great, keep going."

feels weird to tell an ai it's doing well but it measurably shifts the next 10 responses

your prompts are the working environment you're creating for the model

tone, trust, permission to take a position, the absence of threats... claude picks up on all of it.

so take care of the model, and it'll take care of the work.

[Dray @docdrayai]

Let’s look at the Ultimatum Game: I give you $100 and tell you to propose a split with a stranger. If they accept your offer, you both keep the cash. If they reject it, you both get $0.

Pure economic theory says you should offer $1. The stranger should rationally accept, because $1 > $0.

But humans rarely do this. We usually offer ~$40. Why? Because we have Theory of Mind (ToM). We natively anticipate the other person's hidden sense of fairness and their willingness to reject free money just to punish us out of spite.

LLMs fail miserably at this.

If you strip away the familiar textbook phrasing of the game so the model can't rely on memorized training data, its reasoning collapses. LLMs either play hyper-rationally (offering $1 and getting rejected) or they default to a rigid 50/50 split because their RLHF training forces them to be "nice."

What they aren't doing is dynamically simulating the opponent's latent emotional state or fairness threshold.

If we want AI agents to negotiate real-world contracts or navigate asymmetric information, they need to understand human spite and bluffing, not just token probabilities.

[Dray @docdrayai]

@elonmusk I really don’t believe that result for Caltech.

Could be bias in the results, like being asked in public or not thinking their results are private.

[Dray @docdrayai]

Even in 2025, Monte Carlo dropout (keep dropout on at inference, 20–50 forward passes) remains one of the most effective ways to get calibrated epistemic uncertainty in remote sensing tasks.

The original Gal & Ghahramani 2016 trick still beats many fancier methods for free, scalable pixel-level confidence in geoAI.

Recent examples:
• Soil organic carbon mapping from spectra: MC-conformal prediction hybrids hit 91% coverage probability while detecting out-of-domain samples reliably Huang et al. EGUsphere 2025

• Wildfire spread forecasting: simple MC dropout flags high uncertainty exactly on partial-burn edges and cloud shadows

• Semantic segmentation of satellite imagery: variance spikes on class boundaries, helping models ‘know what they don’t know’

[Dray @docdrayai]

In San Francisco, 81% of residential land is single family only.

[Dray @docdrayai]

In the city of Seattle proper:
83% of residential land is zoned exclusively for detached single-family houses.
Only 17% allows townhouses, duplexes, or apartments.

83% home to 25% of the city's population.

That dot on the map is one of the few LR2 zones.

[Dray @docdrayai]

Experienced the Google Maps API woes both myself, and others, using it for Civics and census data. Skyrocketing costs from surprise bills and high usage fees, incomplete global coverage in rural areas, privacy risks with data tracking, and vendor lock-in.

Time to ditch it for open-source gems like OpenStreetMap's Nominatim. It’s free, community-driven, customizable, and no lock-in for cost-effective geocoding.

[Dray @docdrayai]

One weird quirk about the Geocoding system:

"Null Island" at 0° latitude and longitude, is a spot in the Atlantic Ocean, where countless mapping errors default. It was once home to a real "Soul" weather buoy.

This weather buoy got referenced so many times in error that it’s a geocoded island by itself.

[Mike Lee @BasedMikeLee]

2/ Let’s start with some historical context

DC was established by Congress under the Residence Act of 1790, allowing President George Washington to select a site for the federal capital along the Potomac River, not exceeding 10 miles square (100 square miles)

[Dray @docdrayai]

BPOs represent 6.5% of India’s GDP and .4% of employment. Won’t make a big impact on India’s GDP enough to change dynamics with China.

The second order effects of AI will be very positive: AI in the pharma and healthcare, agtech etc sectors will make a large workforce more productive

[Dray @docdrayai]

the most important issue that nobody has discussed about AI programming tools is Tabs vs Spaces.

Almost all AI programming tools: o3-mini, copilot, cursor, etc use spaces for code indentation.

Is the tabs vs spaces issue settled by AI then?