Paul Roberts

An LLM is an error prone probabilistic engine wrapped in a while loop. It is not intelligent. An engineering problem. Not first contact with a new life form.

[Paul Roberts @gammyKneeUK]

@nickkokonas Do you know what an LLM is? Are you aware of the hard ceiling - their limitations? Or are you just guzzling down the kool aid without engaging your brain. Apply some logical reasoning to the claims that you're accepting in blind faith.

[Paul Roberts @gammyKneeUK]

@egosublime @rcpsych It is for me hunter. I would struggle to hold down a job, but self employed and able to live my life my way - I wouldn't swap ADHD cognitive style for the world. The biggest drawback is not being able to apply myself to anything unless I'm interested - and then I'm obsessed! LOL

[Paul Roberts @gammyKneeUK]

@cb_doge Grok isn't intelligent. It's not capable of driving a robot or a self driving car. And never will be. It can only perform well defined tasks. Poor Elon mistook user intelligence steering the AI for emergent intelligence in the AI. He got excited over an illusion and overpromised.

[Paul Roberts @gammyKneeUK]

@FLI_org @AnthonyNAguirre @NEWSMAX Nope. It's not smart. It's just a dumb probabilistic engine wrapped in a while loop with poorly defined goals. Had they applied engineering principles instead of treating it like first contact with an alien lifeform 'agents' would be banned and those responsible for them liable.

[Paul Roberts @gammyKneeUK]

@Polymarket The West has certainly lost it. Japanese, Russian and Chinese AI leaves it in the dust.

[Paul Roberts @gammyKneeUK]

@Polymarket No one is going to achieve AGI. Certainly not with LLMs. It's just hype to part fools with their cash.

[Paul Roberts @gammyKneeUK]

@techday_au How can decisions made by inherently error prone probability engines ever be remotely reliable. Does anyone in AI governance have their brain engaged?

[Paul Roberts @gammyKneeUK]

@ABridgen I suspect wrongful GCHQ surveillance and have been met with a 16 month wall of silence, SAR's refusals etc. I have done nothing wrong - it seems related to battles over mining rights for cognitive data via AI. Any advice? I'm hitting a brick wall for answers.

[Paul Roberts @gammyKneeUK]

@ABridgen Andrew, I sent a SAR to my MP in December 2025. He received it on 17th August 2026. With the date modified. I have a digital file with a cryptographically matching counterpart in Parliament's systems, but the core metadata (Date:) has been surgically altered in transit.

[Paul Roberts @gammyKneeUK]

@omgsidewalks The climate always changes. What isn't scientifically proven is the impact that humans are having - for better or worse.

[Paul Roberts @gammyKneeUK]

@APMI93301887 @StevenScott321 @goldman It works *sometimes*. Lulling people into a false sense of security. There is no intelligence behind it.

[Paul Roberts @gammyKneeUK]

@APMI93301887 @StevenScott321 @goldman It is untrained weights with a training cutoff date later than that of the qwen models you are using. I am making the point that I have steered untrained weights to do high level math - I fully understand how an LLM works. And it's limitations.

[Paul Roberts @gammyKneeUK]

@KanishkaNarayan Distinctly British AI’ is like claiming ‘British gravity.’ Maths doesn’t have a passport.

The growth is the sound of billions being poured into a hole, hoping a magic beanstalk grows out of it. (hint - it ain't going to happen. Dot-com anyone?).

[Paul Roberts @gammyKneeUK]

@TivadarDanka Yes. They can be steered and made useful. But when there's no user steering it's pretty much pot luck. Chinese models are better. Russian too. Western AI is slop due to the extreme narrative control and thick 'safety' layers. Dangerous.

[Paul Roberts @gammyKneeUK]

@sethlazar It's more like the establishment freaking out because they can't get the LLMs to go along with their bat shit crazy, un-tethered from reality, narrative.

It's narrative control. Or rather - the panic is - they've lost control of the narrative.

[Paul Roberts @gammyKneeUK]

@APMI93301887 @StevenScott321 @goldman The danger is, due to dishonest marketing and output that looks right and sometimes is right - people are trusting AI that is effectively autocomplete.

[Paul Roberts @gammyKneeUK]

@APMI93301887 @StevenScott321 @goldman Qwen 3.5 35B Guardrails stripped:

"When I generate text directly (without external tools like Python or Wolfram Alpha), my raw token-prediction error rate for complex math typically sits around 15–25%. This means that roughly 1 in 4 to 1 in 5 outputs will contain an error."

[Paul Roberts @gammyKneeUK]

@APMI93301887 @StevenScott321 @goldman Qwen 3.5 35B Guardrails stripped:

"With tools - Complex Math (Calculus, Linear Algebra): Moderate reliability (~70–80%)"

[Paul Roberts @gammyKneeUK]

@LuizaJarovsky Too late. The horse has bolted. It's in every back bedroom and basement - with the guardrails stripped out. Meanwhile - regulators are still refining their first contact protocols, keen to make a good impression with a new alien lifeform. Clown world.