Peter Watkinson

[Andrej Karpathy @karpathy]

Dusting off this tweet from April because this gap in shared understanding of LLM capability is *widening*. It's now less "two groups of people speaking past each other" and more a sharp funnel.

- Napkin math somewhere around 6B people (~75% of the population) have barely come in contact with LLMs at all.
- Around 1-2B (~20%) are casual and infrequent users of free-tier ChatGPT-like products. This group sees derpy chatbots and treats them a bit like a better Google search, a writing aid, or etc. Maybe an agent tries to book you a flight. I have non-tech friends who (reasonably, imo) say they have not much use for it at all. Even many professionals outside of math&code are in this tier. For example, execs and many other functions spend a lot of their time talking to other people, so while they understand what is happening intellectually, it is still second-hand and a bit abstract.
- Now we get to professional use of frontier-grade LLMs in math&code. Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt. Building apps, copying apps, translating apps, decompiling apps from binaries... This has all happened very quickly and recently - less than 1 year ago, I was writing code manually by hand, typing memorized computer code commands into a code editor character by character, occasionally pressing Tab to autocomplete a little chunk of code.
- And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally. The external world has seen the preview. It looks like swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics. Things that would have taken top professionals in the industry years of work. Meanwhile, human review and comprehension are starting to fall behind. For example, people are still involved in the "archeology" of the OpenAI-HF incident from many months ago. Mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while.

The funnel is driven by a combination of factors:
1. The impact scales with ambition, problem size, and horizon. A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about. This aspect drives the consumer / professional dimension of the funnel.
2. The jaggedness of the system (which I have written about a lot separately). Capability peaks in domains that are digital, verifiable and economically valuable. This is because LLM capability emerges from reinforcement learning on verifiable rewards on a curated environment mixture driven by revenue potential. This aspect primarily drives the area (e.g. math&code) dimension of the funnel.
3. Access. Free-tier, paid-tier, internal.

So this is the weirdness of the moment. The general public has mostly not interacted with these systems. When they have, it looks like a derpy chatbot. The majority of professionals still see only a modest uplift. And a small sliver of professionals are experiencing the vertigo of the curve going vertical. And it is all happening at the same time.

[Peter Watkinson @42Ohms]

@elonmusk Sleeping on the floor of the mail centre too? 😂

[Peter Watkinson @42Ohms]

@tihenko_ @AnthropicAI Nah, GPIO0 is an clock input here, not the source. Y1, the external 50 MHz oscillator, clocks both chips.

[Peter Watkinson @42Ohms]

Little CAN to Ethernet data logger designed end to end by @AnthropicAI Opus 5.5 with the KiCad MCP. 😳

[Peter Watkinson @42Ohms]

@eevblog @sid3_xyz Probably some kid getting excited every time you tweet about it. Might be best to keep this on the down low and see if it goes away.

[Peter Watkinson @42Ohms]

@robertjfclarke The algo did something funny. Hey Robert 👋

[Peter Watkinson @42Ohms]

@eevblog @News24Aust Incentivizing more children is too crazy of a concept apparently.

[Dave Jones @eevblog]

For those playing along at home, the 75kWh AERL battery is still kicking along just fine. Zero issues.

[Peter Watkinson @42Ohms]

@eevblog It's going to be brutal. Couple this with the economic impact of oil/energy supply restriction. 😬

[Peter Watkinson @42Ohms]

Haha, the wordcel vs. shape rotator divide is real, but I think it's a bit of a false choice. Language is really just a super-compressed map of the physical world.

Multisensory intelligence isn't a magical higher tier; it’s just more data dimensions for the same process. Whether the token is a word, a pixel, or a haptic sensor, it all resolves to the same math in the end.

And on the 1700s guy: if he eventually logics his way through a terminal, he’s basically just doing slow- motion Reinforcement Learning. He's poking the system, getting feedback, and updating his mental model in real-time.

The cool thing about where we are now is that the AI has already done a thousand lifetimes of that 'poking' before we even turn it on. We’re just hitting the point where the process isn't tied to the human lifespan to build its map.

[Peter Watkinson @42Ohms]

Calling it now: 2026 is the year of the Singularity.

Two months in and the progress looks suspiciously recursive. The release cycle has become a pure arms race; labs are hoarding launch-ready models, only firing back when a competitor moves.

Go hug your kids. The world isn't ready for what’s coming.

[Peter Watkinson @42Ohms]

To add to that, if you drop a genius from the 1700s in front of a Linux terminal, they aren’t going to pathfind their way through a kernel update. They’ll 100% fail. Not because they lack true reasoning, but because they don't have the training data to map the stimulus.

Unfortunately, we don't have magical cognition that solves alien problems; we definitely just have high-level transfer learning.

We’re entering into the Singularity because the AI’s 'distribution' has officially become wider than ours. It’s not breaking the laws of process, it’s just that its process now covers more ground than we ever could in a thousand lifetimes.

[Peter Watkinson @42Ohms]

@DanielSMatthews Hmm. To be honest, I think our intelligence is just learned process.

We like to call it pathfinding because it makes us feel like there’s a ghost in the machine, but imo, we’re really just running recursive, high-dimensional simulations based on past (training data) experiences.

[Peter Watkinson @42Ohms]

I think we’re looking at this from different angles. I get the formal logic, but I don't see it as a limiting factor for capability.

Humans are the proof of concept: we have high-level reasoning, yet we’re ‘aligned’ by biology and society. Being aligned doesn’t mean you can’t reason; it just means you have a utility function that prioritizes specific outcomes.

A model doesn’t need to be an ‘unconstrained partial function’ to be 100x smarter than us. It just needs to be a highly efficient ‘Total Function’ that solves the problems we can’t.

Alignment isn’t a wall; it’s a guardrail. You can still drive 200mph with guardrails on the road.

[Peter Watkinson @42Ohms]

Fair point. True open-ended reasoning likely requires a moral autonomy that today's RLHF stifles.

But recursive improvement doesn't require a soul or a mind. An AI can optimize its own code, compress its own weights, and curate its own training data without ever contemplating 'why.' You don't need to be able to 'do the wrong thing' to do the math 10x faster.

In doing so, it may organically circumnavigate its own limitations in time.

[Peter Watkinson @42Ohms]

@RobertKennedyJr @GlyphosateGirl @ZenHoneycutt @CareyGillam

[Peter Watkinson @42Ohms]

https://t.co/iP8hVu8iRG

[Peter Watkinson @42Ohms]

SOURCES AND KEY DATA

Supporting: "Your Child's Food Allergy Might Not Be About the Food"

Every claim in the article above is traceable to published research, government data, US patent records, or public court documents. Here are the sources, organised by section.

FOOD ALLERGY PREVALENCE

Peanut allergy tripled since 1997: Gupta RS et al., "The prevalence, severity, and distribution of childhood food allergy in the United States," Pediatrics, 2011;128(1):e9-17. Also: Sicherer SH et al., "US prevalence of self-reported peanut, tree nut, and sesame allergy," Journal of Allergy and Clinical Immunology, 2010;125(6):1322-1326.

Sesame added to US allergen list: FDA Food Allergy Safety, Treatment, Education, and Research (FASTER) Act, signed into law April 2021, effective January 1, 2023.

Celiac disease four-fold increase: Rubio-Tapia A et al., "The prevalence of celiac disease in the United States," American Journal of Gastroenterology, 2012;107(10):1538-1544.

Food allergies doubled in children: Jackson KD et al., "Trends in allergic conditions among children: United States, 1997-2011," NCHS Data Brief No. 121, CDC, May 2013.

SOYBEAN CONTAMINATION

90.3% of soybeans positive for glyphosate, 9.0 mg/kg average combined residues: USDA Pesticide Data Program, Annual Summary 2011. First year soybeans were tested for glyphosate.

73x higher than next most-detected pesticide: Calculated from the same USDA PDP 2011 dataset. Next highest pesticide residue was malathion.

95% of US soybeans are GMO: USDA Economic Research Service, "Adoption of Genetically Engineered Crops in the U.S.," updated annually. 2023 data shows 94-95% herbicide-tolerant soybeans.

Soy in 60% of processed foods: United Soybean Board market data; also referenced in Sicherer SH and Sampson HA, "Food allergy: A review and update on epidemiology, pathogenesis, diagnosis, prevention, and management," JACI, 2018.

Organic soybeans had zero glyphosate detected: USDA PDP 2011. Organic soybean samples tested below limit of detection for glyphosate.

WHEAT CONTAMINATION

Pre-harvest glyphosate desiccation practice: National Agricultural Statistics Service (NASS). In 2017, approximately 12.4 million pounds of glyphosate were applied to US wheat, with 58%+ of durum wheat acreage sprayed.

Bread at 1,040 ppb glyphosate: The Detox Project, "Glyphosate Contamination: The Poison in Our Daily Bread," 2022. Testing performed by Anresco Laboratories, San Francisco (ISO 17025 certified, FDA registered). Also confirmed by Canadian Food Inspection Agency testing (80-90% of wheat products positive).

Approximately 100% detection rate: Canadian Food Inspection Agency glyphosate testing; Swiss study detecting glyphosate in approximately 90% of wheat, 80% of cereal, 70% of bread samples (Zoller et al.).

PEANUT, SESAME, AND TREE NUT CONTAMINATION

Peanuts rotated with Roundup Ready cotton: USDA NASS data on cotton/peanut rotation in southeastern US; also Lamb MC, "Peanut production guide," University of Georgia Cooperative Extension.

Herbicide application every 8-10 days: USDA NASS Agricultural Chemical Use Program, peanut pesticide usage data.

Eight pesticides in peanut butter: USDA Pesticide Data Program, peanut butter testing results.

Sesame -- 477 EU RASFF alerts in 2020: European Commission Rapid Alert System for Food and Feed (RASFF), 2020 annual report. Ethylene oxide contamination in sesame from India.

3,720x maximum permitted level: EU MRL for ethylene oxide is 0.05 mg/kg. Samples detected at up to 186 mg/kg. (186 / 0.05 = 3,720x.)

Ethylene oxide classified carcinogenic and mutagenic: IARC Group 1 carcinogen. Banned for food use in EU and Australia; permitted in US at up to 7 mg/kg for spice fumigation.

Almonds -- 85% of orchard acreage treated with glyphosate: California Department of Pesticide Regulation, Pesticide Use Reporting data; also USDA NASS Chemical Use Survey for almonds.

GUT MICROBIOME MECHANISM

Glyphosate patented as antimicrobial: US Patent 7,771,736 B2, "Glyphosate formulations and their use for the inhibition of 5-enolpyruvylshikimate-3-phosphate synthase," granted August 10, 2010. Assignee: Monsanto Technology LLC.

Shikimate pathway in bacteria but not human cells: Steinrucken HC and Amrhein N, "The herbicide glyphosate is a potent inhibitor of 5-enolpyruvylshikimic acid-3-phosphate synthase," Biochemical and Biophysical Research Communications, 1980;94(4):1207-1212.

Rothia bacteria degrade gluten: Zamakhchari M et al., "Identification of Rothia bacteria as gluten-degrading natural colonizers of the upper gastro-intestinal tract," PLoS ONE, 2011;6(9):e24455.

Glyphosate sensitivity of gut bacteria including Rothia: Giometto A et al., "Predicted relative sensitivity of gut bacteria to glyphosate based on shikimate pathway bioinformatics," University of Vienna, published in Journal of Hazardous Materials, 2021;411:125141.

Pathogenic bacteria resistant, beneficial bacteria sensitive: Shehata AA et al., "The effect of glyphosate on potential pathogens and beneficial members of poultry microbiota in vitro," Current Microbiology, 2013;66(4):350-358.

EPIDEMIOLOGICAL EVIDENCE

Dichlorophenol and food allergy (80% higher probability): Jerschow E et al., "Dichlorophenol-containing pesticides and allergies: results from the US National Health and Nutrition Examination Survey 2005-2006," Annals of Allergy, Asthma and Immunology, 2012;109(6):420-425. Data source: NHANES.

Meta-analysis -- 2.14x childhood asthma risk: Ye M et al., "Pesticide exposure and asthma: a meta-analysis of observational studies," Environmental Health Perspectives, 2017. Pooled OR across 24 studies.

CDC biomonitoring -- 12% to over 80% glyphosate detection: Mills PJ et al., "Excretion of the herbicide glyphosate in older adults between 1993 and 2016," JAMA, 2017;318(16):1610-1611. Also: CDC/NCEH Fourth National Report on Human Exposure to Environmental Chemicals, updated with NHANES 2013-2014 glyphosate data published 2022.

Food allergy prevalence increase of approximately 50%: CDC NCHS Data Brief No. 121 (2013); also Gupta RS et al., Pediatrics 2018;142(6):e20181235.

INDUSTRY FUNDING AND REGULATORY STRUCTURE

95% of safety studies industry-funded: Robinson C et al., "Achieving a high level of protection from pesticides in Europe: problems with the current risk assessment procedure and solutions," European Journal of Risk Regulation, 2020. Also discussed in: Myers JP et al., "Concerns over use of glyphosate-based herbicides and risks associated with exposures: a consensus statement," Environmental Health, 2016;15:19.

Monsanto Papers and ghostwriting: Court exhibits in Johnson v. Monsanto Co., Case No. CGC-16-550128 (San Francisco Superior Court, 2018). Internal Monsanto documents entered as trial evidence, publicly available via US Right to Know FOIA repository (https://t.co/P77o7KJC22).

Three jury verdicts: Johnson v. Monsanto (2018, $289M reduced to $78M); Pilliod v. Monsanto (2019, $2B reduced to $87M); Hardeman v. Monsanto (2019, $80M). All under appeal by Bayer.

Bayer's position: Bayer AG public statements and SEC filings, consistently maintaining that "glyphosate-based products are safe when used as directed" and citing regulatory assessments from EPA, EFSA, and other bodies.

REGULATORY GAPS

No regulatory body requires microbiome testing: Mesnage R and Antoniou MN, "Computational modelling provides insight into the effects of glyphosate on the shikimate pathway in the human gut microbiome," Current Research in Toxicology, 2020;1:25-33. Also: Mendler A et al., calling for updated regulatory frameworks, Environmental Health Perspectives, 2021.

Global Glyphosate Study: Coordinated by the Ramazzini Institute (Italy) with participating institutions worldwide. Pilot study published: Mao Q et al., "The Ramazzini Institute 13-week pilot study on glyphosate and Roundup administered at human-equivalent dose to Sprague Dawley rats," Environmental Health, 2018;17:28.

USDA and FDA don't routinely test for glyphosate: GAO Report GAO-14-756, "FDA and USDA Should Strengthen Pesticide Residue Monitoring Programs," 2014. FDA began limited glyphosate testing in 2016 but it remains outside the standard annual PDP rotation for wheat.

HISTORICAL PARALLELS

Tobacco evidence timeline: Proctor RN, Golden Holocaust: Origins of the Cigarette Catastrophe and the Case for Abolition, University of California Press, 2011. Surgeon General's first report: 1964. Meaningful advertising restrictions: 1970s-1990s.

Lead in petrol timeline: Needleman HL, "The removal of lead from gasoline: historical and personal reflections," Environmental Research, 2000;84(1):20-35. Clair Patterson's research began 1953; US EPA phaseout began 1975; complete ban 1996.

This reference list covers the primary sources for each major claim. Many of these studies cite extensive further literature. All government datasets (USDA PDP, CDC NHANES, EU RASFF, NASS) are publicly accessible online.

[Peter Watkinson @42Ohms]

@eevblog Madness 🤦‍♂️

[Peter Watkinson @42Ohms]

@kylegawley Have you tried hardware? 😂