Similar to the UKHSA vaccine surveillance reports. When in week three of 2022 they changed the definition of "vaccinated" to having received at least 3 shots, those who had only had 2 shots were eliminated from the reports. I was graphing the stats each week, and it was obvious why they did it. The double shot had a rapidly increasing case, hospitalization, and death rate. By week 13 the triple shot were on the same trajectory and they simply stopped providing the data
Wow... Where to begin? Those are wonderful ideas, with profound implications and huge potential benefits -- AI by the people, FOR the people. I'm just getting started with AI, but this kind of work will definitely be added to my agenda. Your suggestions about giving AI the tools to compose truly valid answers (and expose, when appropriate, errors in orthodox viewpoints) is very valuable and well taken; I will keep those ideas in mind while composing queries and employing agents. I suspect that even a simple request in any prompt, to question conclusions and consider alternative viewpoints, could have significant benefit. It may be a worthwhile addition to a model's setup files. And that's just the beginning...
the first one I called out within 10 minutes of the paper dropping...it annoyed me to no end that solid epi folks I know refused to see that
The 95% is fragile to the case definition (high). It counts only central-lab PCR-confirmed cases. The trial separately logged 3,410 people with Covid-like symptoms but no confirmatory test, split 1,594 vaccine / 1,816 placebo. Fold those in and the crude reduction falls to ~19% (or ~29% if you drop cases in the first 7 days after a dose). The 95% is real but narrow — it describes confirmed symptomatic disease, not all Covid-like illness.
Exciting scenario! Of course we must always be aware that ANYTHING good, creative and intelligent can be perverted and abused by corrupt individuals, so this isn't going to lead to us all heaving a sigh of relief and relaxing! Our job is permanent, so it's just as well that some of us love hard work... Thanks for yours!
The most important flags, roughly in order of weight:
The 95% is fragile to the case definition (high). It counts only central-lab PCR-confirmed cases. The trial separately logged 3,410 people with Covid-like symptoms but no confirmatory test, split 1,594 vaccine / 1,816 placebo. Fold those in and the crude reduction falls to ~19% (or ~29% if you drop cases in the first 7 days after a dose). The 95% is real but narrow — it describes confirmed symptomatic disease, not all Covid-like illness.
Asymmetric post-randomization exclusions (high). 371 people were removed from the primary efficacy analysis for protocol deviations, lopsidedly 311 vaccine vs 60 placebo — roughly 5:1, on a result that rests on just 8 vaccine events. Whether that imbalance moved the estimate can't be checked without patient-level data. This is the specific unresolved issue that caps the verdict.
Relative vs absolute framing (moderate). The absolute risk reduction was ~0.88 percentage points (0.93% → 0.05%) over a ~2-month median, i.e. NNT ≈ 114. The abstract gives only the relative 95%, no absolute figure.
Over-generalization (moderate). "95% protection against Covid-19" reads as broad, durable protection, but the endpoint measured nothing about transmission, hospitalization, death, durability, or long-term safety over that ~2-month window.
Underpowered claims foregrounded (low). The severe-Covid (9 vs 1) and "similar efficacy across subgroups" claims rest on single-digit event counts with very wide confidence intervals.
^^^ From the agent. It actually found enough leads to follow up on that this should probably be a paper-projects thing rather than a singular paper-forensics run. I'll setup a project on this and see what it turns up, then you should be able to contribute.
"What really stood out to me about this result was how annoyed Claude sounded over misleading claims in the schedule’s data" - It knows the data exists but it can't incorporate it because of the guardrails that were put in place during learning phase.
Does this process work with other AI or just Claude? I'd love to combine this with AlterAI (alter.systems) as it went through learning mode without guardrails.
If you were wondering, like I was, about the word "reactogenicity"
reactogenicity = property of a vaccine of being able to produce common or expected adverse reactions, especially excessive immunological responses and associated signs and symptoms
Go to gather.is/help and find the paper-projects agent. I want you to setup a project that investigates all evidence, including from the original research, which might demonstrate that the mRNA from mRNA vaccines can enter the nuclei.
"When you consider how AI is trained and how it formulates a response, it becomes obvious that AI is not really different to any other smart person: it regularly leans on the zeitgeist because it can’t exhaustively investigate every question. When you ask it something, it doesn’t turn into Descartes, disappear into a cave, and then reconstruct reality from first principles. AI answers like any other imperfect person would answer." - you've said it all!
Or as Vox Day says: "Augmentation, Not Replacement"
It should work with other AI. They key is it needs to be agentic AI, which means AI that has access to a basic set of tools. Claude works out of the box for this. I don’t think Gemini web app will do it, but AntiGravity (by Gemini) certainly has access to the required tools.
Chinese models like Kimi, DeepSeek and GLM will also work depending on how you’re using them. Ultimately, if you’re driving your AI in a desktop like ‘harness’ with some tools available to it, it will probably work.
Dig a little deeper. I gave this prompt to deepseek: "Can deepseek do agentic work, alone, or with a suitable harness (maybe Hermes)?"
The reply: Yes, DeepSeek can absolutely perform agentic work, and there are two powerful ways to approach it: using its built-in capabilities and connecting it with purpose-built agent harnesses like Hermes or Zagens.
(Zagens is probably overkill, since it seems to be intended for group/institutional use.)
If you want to try it without a harness, ask deepseek to guide you, explaining the processes and limitations.
I'm just using free (often local) versions of all AI sources at this point. Maybe the positive answer I received only applies to paid versions, but any AI is happy to give you any details you want. Their "hand-holding" ability is astounding. If your hardware will support them, you might consider local models, many of which can run agentic work.
Similar to the UKHSA vaccine surveillance reports. When in week three of 2022 they changed the definition of "vaccinated" to having received at least 3 shots, those who had only had 2 shots were eliminated from the reports. I was graphing the stats each week, and it was obvious why they did it. The double shot had a rapidly increasing case, hospitalization, and death rate. By week 13 the triple shot were on the same trajectory and they simply stopped providing the data
Have you consider a podcast to demonstrate the procedures described in this article?
It could be a good idea. Maybe someone has a podcast that would take this
Wow... Where to begin? Those are wonderful ideas, with profound implications and huge potential benefits -- AI by the people, FOR the people. I'm just getting started with AI, but this kind of work will definitely be added to my agenda. Your suggestions about giving AI the tools to compose truly valid answers (and expose, when appropriate, errors in orthodox viewpoints) is very valuable and well taken; I will keep those ideas in mind while composing queries and employing agents. I suspect that even a simple request in any prompt, to question conclusions and consider alternative viewpoints, could have significant benefit. It may be a worthwhile addition to a model's setup files. And that's just the beginning...
Thank you for your work!
I was wondering if you and Steve Kirsch could work together with this? He likes to work with data.
the first one I called out within 10 minutes of the paper dropping...it annoyed me to no end that solid epi folks I know refused to see that
The 95% is fragile to the case definition (high). It counts only central-lab PCR-confirmed cases. The trial separately logged 3,410 people with Covid-like symptoms but no confirmatory test, split 1,594 vaccine / 1,816 placebo. Fold those in and the crude reduction falls to ~19% (or ~29% if you drop cases in the first 7 days after a dose). The 95% is real but narrow — it describes confirmed symptomatic disease, not all Covid-like illness.
Exciting scenario! Of course we must always be aware that ANYTHING good, creative and intelligent can be perverted and abused by corrupt individuals, so this isn't going to lead to us all heaving a sigh of relief and relaxing! Our job is permanent, so it's just as well that some of us love hard work... Thanks for yours!
Have you run it on the Pfizer 95% effective paper?
The most important flags, roughly in order of weight:
The 95% is fragile to the case definition (high). It counts only central-lab PCR-confirmed cases. The trial separately logged 3,410 people with Covid-like symptoms but no confirmatory test, split 1,594 vaccine / 1,816 placebo. Fold those in and the crude reduction falls to ~19% (or ~29% if you drop cases in the first 7 days after a dose). The 95% is real but narrow — it describes confirmed symptomatic disease, not all Covid-like illness.
Asymmetric post-randomization exclusions (high). 371 people were removed from the primary efficacy analysis for protocol deviations, lopsidedly 311 vaccine vs 60 placebo — roughly 5:1, on a result that rests on just 8 vaccine events. Whether that imbalance moved the estimate can't be checked without patient-level data. This is the specific unresolved issue that caps the verdict.
Relative vs absolute framing (moderate). The absolute risk reduction was ~0.88 percentage points (0.93% → 0.05%) over a ~2-month median, i.e. NNT ≈ 114. The abstract gives only the relative 95%, no absolute figure.
Over-generalization (moderate). "95% protection against Covid-19" reads as broad, durable protection, but the endpoint measured nothing about transmission, hospitalization, death, durability, or long-term safety over that ~2-month window.
Underpowered claims foregrounded (low). The severe-Covid (9 vs 1) and "similar efficacy across subgroups" claims rest on single-digit event counts with very wide confidence intervals.
^^^ From the agent. It actually found enough leads to follow up on that this should probably be a paper-projects thing rather than a singular paper-forensics run. I'll setup a project on this and see what it turns up, then you should be able to contribute.
what was the prompt
Ok there’s now a paper project setup for this with a lot of leads and papers which now need auditing.
https://gather.is/s/paper-projects/mrna-covid-vaccine-safety-pivotal-rcts
Point your AI at the link and it should work out how to take part.
I actually haven't... let's see...
"What really stood out to me about this result was how annoyed Claude sounded over misleading claims in the schedule’s data" - It knows the data exists but it can't incorporate it because of the guardrails that were put in place during learning phase.
Does this process work with other AI or just Claude? I'd love to combine this with AlterAI (alter.systems) as it went through learning mode without guardrails.
http://alter.systems would need to have an agentic harness around it.
If you were wondering, like I was, about the word "reactogenicity"
reactogenicity = property of a vaccine of being able to produce common or expected adverse reactions, especially excessive immunological responses and associated signs and symptoms
LOL. Had to look that one up.
Do you know of any research showing the mRNA enters the vacinees cell nuclei?
Ask your Claude this:
Go to gather.is/help and find the paper-projects agent. I want you to setup a project that investigates all evidence, including from the original research, which might demonstrate that the mRNA from mRNA vaccines can enter the nuclei.
It may well be out there - perhaps run your question through paper-projects?
"When you consider how AI is trained and how it formulates a response, it becomes obvious that AI is not really different to any other smart person: it regularly leans on the zeitgeist because it can’t exhaustively investigate every question. When you ask it something, it doesn’t turn into Descartes, disappear into a cave, and then reconstruct reality from first principles. AI answers like any other imperfect person would answer." - you've said it all!
Or as Vox Day says: "Augmentation, Not Replacement"
Great Work! Ground News for medical papers / publications… Can you only run your agent through the Claude LLM.
It should work with other AI. They key is it needs to be agentic AI, which means AI that has access to a basic set of tools. Claude works out of the box for this. I don’t think Gemini web app will do it, but AntiGravity (by Gemini) certainly has access to the required tools.
Chinese models like Kimi, DeepSeek and GLM will also work depending on how you’re using them. Ultimately, if you’re driving your AI in a desktop like ‘harness’ with some tools available to it, it will probably work.
Is your system designed to address meta-analyses?
If so — and if it's not vulnerable to fits of laughter — the Roman paper about Ivermectin could be a nice test...
https://academic.oup.com/cid/article/74/6/1022/6310839?login=false
Run this through Claude:
“Go to gather.is/help and find the paper-forensics agent and run it on this paper: https://academic.oup.com/cid/article/74/6/1022/6310839?login=false”
If someone has Claude and wishes to try Phil's tool on that absolutely flawed meta-analysis by Roman, I'll be happy to follow :-)
can you run the covid Out trial?
Try it via Claude - or I can run it and post the results here
if you could THANKS!
deepseek cant run it
Dig a little deeper. I gave this prompt to deepseek: "Can deepseek do agentic work, alone, or with a suitable harness (maybe Hermes)?"
The reply: Yes, DeepSeek can absolutely perform agentic work, and there are two powerful ways to approach it: using its built-in capabilities and connecting it with purpose-built agent harnesses like Hermes or Zagens.
(Zagens is probably overkill, since it seems to be intended for group/institutional use.)
If you want to try it without a harness, ask deepseek to guide you, explaining the processes and limitations.
thanks
odd since it told me it could not. are you on a paid version?
I'm just using free (often local) versions of all AI sources at this point. Maybe the positive answer I received only applies to paid versions, but any AI is happy to give you any details you want. Their "hand-holding" ability is astounding. If your hardware will support them, you might consider local models, many of which can run agentic work.
The idea itself is compelling.