Message from Charles Aulds

What a surprise:  to learn too late that the United States is not only a poor neighbor, but an untrustworthy ally of Canada. The US has become active antagonist to Canada. An enemy. And not for the first time.

I was surprised to learn (and we were never taught this in school) that President Abraham Lincoln's Secretary of State, William Henry Seward, was determined to take Canada, one century and a half ago. Seward was a rabid expansionist. A true believer in America's "Manifest destiny."

It was Seward who negotiated the US purchase of the Danish West Indies in 1865. Today that's known as the US Virgin Islands.

It was Seward who bought Alaska from Russia two years later, in 1867.

It was Seward who tried to buy Greenland that same year (1867) from Denmark.

Seward wanted Greenland because (he said) that Canada, sandwiched between two large American states, would then inevitably become part of Canada. Canadians would have no other choice than to submit to US annexation.

It was in 1867 that the Province of Canada (what is now Ontario and Québec) and two British North American colonies (New Brunswick, Nova Scotia) united and became a sovereign nation. The year of Confederation (1867) is significant; it was not the result of coincidental timing.

Will Americans have a second civil war? Who really cares? This MAGA shit gotta stop sometime.

___

Charles

Reports on using light to implement AI operations in order to save energy

 

Photonic (light-based) AI computing has rapidly emerged as one of the most promising approaches to dramatically cutting the enormous energy costs of AI systems. Multiple research groups and companies are now demonstrating working prototypes across a range of applications.

Why Light Instead of Electrons?

Photons — the particles of light — don't interact with each other under normal conditions, meaning many light signals can pass through the same system simultaneously, processing large data sets at the speed of light with very low latency. Conventional electronic AI hardware loses massive amounts of energy to resistance and heat in transistors, but optical systems can perform the same mathematical operations — particularly the heavy matrix multiplications that underpin neural networks — with far less power loss. Projections suggest optical accelerators could cut AI energy use by up to 90% compared to electronic equivalents.lumai+2

Key Research Breakthroughs

Several major advances have been announced in rapid succession:

• MIT fully integrated photonic chip (Dec 2024): MIT researchers built a photonic processor that performs all key deep neural network operations — both linear and nonlinear — entirely in the optical domain, achieving over 96% accuracy in training and computing results in less than half a nanosecond.[physics.mit]​

• University of Florida light-powered chip (Sept 2025): By etching microscopic lenses directly onto silicon, researchers enabled laser-powered computations that cut power use dramatically while maintaining near-perfect accuracy, also demonstrating wavelength multiplexing — running multiple data streams on different colors of light simultaneously.[sciencedaily]​

• UCLA generative AI optical model (Oct 2025): UCLA devised an optical computing strategy that generates novel images and videos using much less energy than conventional generative AI models, published in Nature.[optica-opn]​

• Aalto University single-beam tensor computing (Nov 2025): A method where AI operations occur passively as light travels — requiring no active control or electronic switching — making it compatible with almost any optical platform and extremely low power.[sciencedaily]​

• Penn State "infinity mirror" loop (Feb 2026): A prototype where light is routed through a compact multi-pass optical loop built from everyday LCD and LED components, encoding data directly into light beams and achieving AI inference at dramatically lower energy cost.[psu]​

Where It's Being Applied

The highest-value near-term application is AI inference — the stage where a trained model responds to real-world inputs — which accounts for 80–90% of total AI workload energy. Photonic chips are also being explored for lidar, telecommunications, astronomy, and real-time navigation. Companies like Lightmatter are already commercializing photonic AI accelerators, and Q.ANT has released a photonic AI processor as a standard PCI Express card for integration into existing systems.lightmatter+3

Energy & Sustainability Context

AI data centers are projected to consume as much electricity as an entire country in 2025, with GPUs generating enormous heat that is itself a major operating cost. Photonic computing directly addresses both problems — less electrical power is consumed and far less heat is generated, since light doesn't heat up a medium the way electrical current does through resistance. University of Jena's new research group, funded by the German Federal Ministry with €2.3 million, is taking this further by working on optical computing units as small as the atomic building blocks of crystalline materials — so-called "picophotonic" computing.uni-jena+3

The field is still largely in the prototype and early commercialization phase, with the main challenge being full integration of all AI operations onto a single photonic chip and scaling to production volumes.[pmc.ncbi.nlm.nih]​

AI comparisons

Here’s a concise, “all‑axes” comparison of the current top‑tier stacks: Gemini (2.5 Pro / 3 Pro), GPT (GPT‑4.1 / ChatGPT‑5 family), and Claude (Claude 4 Sonnet/Opus). I’ll focus on what actually differs in practice.openai+4

High‑level positioning

• Gemini: Multimodal/context monster with tight Google integration and massive context; strongest when you mix text, code, images, audio, video or huge corpora.itecsonline+2

• GPT: Most generalist and tool‑rich; very strong reasoning + coding with the broadest native multimodal stack and ecosystem (plugins, agents, third‑party tools).sigmabrowser+3

• Claude: Deliberate thinker optimized for coding, long‑form analysis, safety, and “extended thinking”; strongest on complex software and careful explanations.anthropic+3

Capability table (late‑2025 / early‑2026 snapshot)

Dimension	Gemini top tier (2.5 Pro / 3 Pro)	GPT top tier (GPT‑4.1 / ChatGPT‑5)	Claude top tier (Claude 4 Sonnet / Opus 4.x)
Core strength	Multimodal + huge context + Google integration.datastudios+2	Balanced reasoning, coding, multimodal, tools.openai+2	Best‑in‑class coding and careful reasoning.anthropic+3
Reasoning (general)	Competitive, especially with long or multimodal inputs; slightly behind GPT/Claude on some pure text reasoning benchmarks.itecsonline+1	Very strong general reasoning; 1M‑context “needle in haystack” and graphwalks evals show robust long‑context logic.openai+1	Strongest on many structured reasoning / analysis tasks with “extended thinking” turned on.anthropic+2
Coding	Good–very good; below Claude and often a bit under GPT on SWE‑bench‑style tests.itecsonline+2	Strong, versatile coding and debugging across languages; not the top on SWE‑bench but solid.glbgpt+2	Industry‑leading SWE‑bench Verified (≈72–75%, up to ≈80% with parallel compute).anthropic+4
Multimodal I/O	Native text, images, audio, video in one stack; strong at video and document‑style vision.blog+3	Full unified multimodal (text, image, audio, video) with mature tools; very flexible.openai+2	Primarily text‑centric with maturing image/PDF input; less emphasis on full video/audio pipeline.anthropic+3
Context window	Up to ≈2M tokens on Pro in Vertex / high tiers; Flash also very large.docs.cloud.google+4	Up to 1M tokens in GPT‑4.1 family (API), smaller in standard UI.learn.microsoft+4	Commonly ≈200k in Sonnet/Opus 4; up to 500k in some enterprise tiers.platform.claude+2
Long‑context quality	Designed for massive document/code/video workloads; strong but not always best at fine‑grain reasoning inside huge contexts.datastudios+2	Very good: 1M‑token “needle” and graphwalk benchmarks show robust retrieval in long context.openai+1	Good, but smaller max windows; shines more in careful reasoning than sheer size.platform.claude+1
Speed / latency	Flash models extremely fast (hundreds of tokens/s, sub‑0.3s TTFT); Pro slower but still competitive.deeplearning+2	Balanced; faster than earlier GPT‑4, not as fast as Gemini Flash in most reports.glbgpt+2	Sonnet often mid‑pack for speed; extended‑thinking modes intentionally slower for harder tasks.anthropic+1
Pricing (API ballpark)	Aggressive on context (lots of tokens per dollar); very good value for multimodal + long context.google+2	Generally mid‑range per million tokens; good value given ecosystem and tools.openai+1	Tends to be pricier at the top tiers, but cost justified for orgs that value coding + safety.anthropic+2
Safety / alignment	Strong Google safety stack, guardrails, and filters; conservative on some topics.blog+2	Mainstream OpenAI approach; improved guardrails, but more permissive than Claude on many tasks.openai+2	Most safety‑constrained; Anthropic centers “constitutional AI” and conservative defaults.anthropic+2
Tooling & agents	Deep integration with Google (Workspace, Search, Maps, YouTube) plus Vertex AI; agentic features and Deep Research in Gemini Advanced.blog+3	Rich tool ecosystem: function calling, workflows, agents, plugins; strong third‑party ecosystem.openai+2	Strong for enterprise agents (Bedrock, Vertex, Anthropic API), with emphasis on reliability and governance.anthropic+2
Ecosystem & adoption	Natural choice for Google‑centric orgs; strong in data/Docs/Sheets, Android, and Chrome contexts.gemini+2	Broadest developer and consumer footprint; many libraries, UIs, and SaaS products built around GPT APIs.openai+2	Popular in enterprises that care about risk, compliance, and code quality; embedded in tools like GitHub Copilot agents and Bedrock.anthropic+2

Detailed axes

1. Reasoning and long‑context work

• Gemini: Very strong at multi‑step reasoning when the task uses its strengths—huge context or multimodal inputs—but independent testing often puts it just below Claude for hard coding/logic and roughly on par or slightly under GPT‑4.1 on pure text benchmarks.datastudios+4

• GPT: Excellent generalist; 1M‑token context plus strong “needle‑in‑a‑haystack” and graphwalk scores show it can both hold and use massive context effectively.datacamp+1

• Claude: Often top on structured reasoning and careful analysis, especially when extended‑thinking is enabled, at the cost of latency and price.anthropic+3

If you’re doing long technical reports or philosophical analysis, Claude usually gives the most coherent, reflective write‑ups; GPT is close and more versatile; Gemini is best when those reports must integrate many files, images, or long transcripts.

2. Coding and software engineering

• Benchmarks: Multiple reports put Claude 4 Sonnet/Opus at the top of SWE‑bench‑style benchmarks (≈72–75%+, up to ≈80% with parallel compute), with Gemini 2.5 Pro lagging and GPT‑4.1 in the middle.apidog+4

• Developer workflow:

  • Gemini: Great when your repos live in Google Cloud, and when you need to reason over diagrams, logs, or UI screenshots as part of coding.docs.cloud.google+3

  • GPT: Best “all‑rounder” for IDE integrations, agents, and quick prototypes; wide tooling support (Cursor, VS Code assistants, etc.).glbgpt+3

  • Claude: Favored for refactoring large codebases and debugging subtle issues; popular for “trust it with a big monorepo” tasks.entelligence+4

For your kind of deep, systems‑level coding or analysis, Claude is generally the best first pick; GPT is the second choice for breadth and tools; Gemini is ideal when the code problem is entwined with data, docs, or visual context.

3. Multimodal (text, image, audio, video)

• Gemini: Strong emphasis on multimodality—image, audio, and video understanding plus generation through related models (e.g., Veo, Imagen) in the Google AI Pro/Ultra bundles.blog+4

• GPT: Highly capable unified multimodal model embedded in ChatGPT‑5/4.1, handling images, audio, and video in a single conversational flow, with many creative tools wrapped around it.openai+3

• Claude: Primarily text‑first; image/PDF inputs are supported in many deployments, but it does not yet lean as hard into full video/audio workflows as Gemini/GPT.allmates+3

For drone footage, environmental imagery, and long video or audio logs, Gemini or GPT currently give you more leverage than Claude.

4. Context windows, memory, and “agentic” behaviour

• Context: Gemini often leads on raw context size (up to ≈2M tokens), GPT offers up to 1M, Claude up to 200–500k depending on tier.datastudios+5

• Use of context: GPT shows particularly strong ability to actually retrieve specific content from giant contexts; Claude sacrifices raw size but tends to reason deeply over what it has; Gemini sits between, with emphasis on multimodal long‑context workloads.datacamp+3

• Agents & tools: All three support tool calling and multi‑step workflows; GPT has the most mature ecosystem, Gemini is tightly coupled to Google services, Claude leans into enterprise agents with safety and governance knobs.cloud.google+5

For your “big corpus” hydrology/indigenous‑knowledge research, Gemini and GPT are attractive for sheer context; Claude is ideal when you want slower but more interpretable chains of reasoning over smaller curated chunks.

5. Safety, reliability, and “feel”

• Gemini: Conservative content filters and Google’s safety stack; good for organizations worried about brand risk but sometimes frustrating for edgy or controversial topics.blog+2

• GPT: Middle of the road; stronger safety than early GPT‑4 but not as locked‑down as Claude on some sensitive themes.glbgpt+2

• Claude: Most risk‑averse, with “constitutional AI” and strong refusal behavior; often preferred for regulated sectors and sensitive domains.remio+2

Subjectively: Claude often “feels” like a careful human expert; GPT like a flexible all‑purpose collaborator; Gemini like a powerful multimodal lab instrument wired into Google.

---

If you tell me your main dimensions (e.g., coding vs. research vs. multimedia vs. philosophy) and whether cost/latency matter, I can give a more pointed recommendation like “for X, default to Claude; for Y, switch to Gemini; use GPT when you need Z.”