Punny!
brainchild… I see what you did there :D
Posts by itscomplicated
34 publicly visible posts • joined 16 Mar 2026
Disembodied fruit fly brain joins the crypto speculation swarm
Enjoy the Windows 10 Update Experience at London's O2
Hugging Face is too important to fall into Nvidia's hands
Datacenters face direct hit from China rare earth curbs, as clock runs out on escalated licensing chokeoff
CISA: Most exploited vulnerabilities should have been eradicated decades ago
$1T investment giant Apollo breached after social engineering attack
If you're not using AI to attack your own systems, your adversaries will
Researcher tricks Apple’s Find My into sharing location data with Linux
Microsoft finally fixing File Explorer and the Context Menu
Microsoft tosses Teams Live chat into its feature graveyard
Google Earth's AI makeover survives less than one planetary rotation
Samsung warns memory crunch will last through 2028 as profit rises 19-fold
SpaceX to try its luck again with Starship Flight 13 after engines and weather say no
Flock cameras go up in flames as cops hunt suspected firebugs
Connecting AI agents to outside services explodes the risk radius
Sunday 19th July 2026 19:25 GMT
To be fair, for most of that time software has been deterministic. If you want to limit an API's blast radius you either ensure the calling code doesn't do bad things (from the caller side) or establish permissions on the credential (from the service side)
Neither of those are really viable when you want a (effectively) non-deterministic software system to take action on your behalf automatically. You can't control exactly what it will do and if you lock down the credential too much it won't be able to do what you want. So what do you do? Human-authorize each call? Write some heuristic to monitor the calls? Allow the bad things but have rollback functionality?
It's still being worked on.
Microsoft warns customers AI will mean busier Patch Tuesdays
The AI that spawned MechaHitler and deepfake porn puts on a suit to become legal advisor and Excel jockey
Oracle outlines all the ways it could lose the farm it bet on AI
AI is code – and can't be prompted into being smarter
Sunday 14th June 2026 16:37 GMT
Re: Hilarious article leads to question
There is no difference between text and instructions, it's all just tokens. LLMs are fed input tokens and each one affects the output. There is no reason to expect anything else because this is their design.
The question is how can we prevent things we don't want becoming input to the LLM? The solution is still being worked on.
Sunday 14th June 2026 16:14 GMT
Re: Sanitised Inputs
Exactly. I don't think it's really possible in a single model and session scenario but I'm sure people are working on something like a separate model trained to detect malicious instructions.
It's tricky because the network being affected by the input is how the whole thing works so you can't just instruct the model to disregard certain parts. The agent/harness will have to perform sanitation steps first.
Sunday 14th June 2026 16:08 GMT
Sort of
"AI" isn't just code, is a combination of weights - which is data - and software which uses the weights. While the software and weights determine the peak capability of the system, the prompt determines how close a given session gets to its goal, so in that sense AI can very much be prompted to be "smarter". Dumb prompt, dumb output. Smart prompt, smart output.
It's surprising anyone would expect an LLM not to consider some of the input provided to it. Its whole purpose and method of operation is to take all input and correlate it using the weights. It doing that is not "not smart", it's simply doing what it does. To prevent malicious output one needs to prevent malicious input and this is unfortunately a difficult problem since humans are rather good at encoding malicious intent in language. Even humans are regularly affected by this problem (see: advertising, influencers and scams) so it's hardly surprising AI isn't magically good at it.
KPMG's AI report becomes an accidental demo of AI hallucinations
Friday 12th June 2026 16:23 GMT
Re: "Total Experience: Redefining Excellence in the Age of Agentic AI"
Marketing wonks just can’t help shoving the word EXPERIENCE! into every single thing, even when it doesn’t make sense, as here.
My rule of thumb is when you see the word experience you know you’re reading BS. It certainly holds up here.
Gothenburg's self-driving bus trammed on day one
Google explains how it will infuse ads into AI answers
VMware quietly debuts Arm hypervisor tech preview
LocalSend puts your sneakernet out of business
Qualcomm teases ‘dedicated CPU for agentic experiences’ and ‘agentic smartphones’
Fewer users, fatter wallets is why Anthropic tops OpenAI in LLM revenue stakes
Apple has an opportunity to rediscover humanity in its push toward AI
Wednesday 22nd April 2026 08:59 GMT
Hear, hear! Apple regularly states (including in Tim's Community Letter) that they put customers at the center of everything they do. To an extent they do but there is an unfortunate list of things where they clearly do not.
Now they have a great opportunity to undo some of those things. I hope they avail themselves of it.
NASA sets 'impossible' ground rules for relocation of 'flown space vehicle'
ChatGPT advised exec on how to fire Subnautica founders to avoid payout, court ruling says
Nvidia's DLSS 5 promises to bring you out the other side of the uncanny valley
Monday 16th March 2026 23:04 GMT
Performance
Beyond the visual quality improvements what I find exciting about this is how it might impact performance.
In the same way upscalers render at a lower resolution, neural rendering might allow the engine to reduce quality settings, using much simpler geometry, textures and shaders and eschewing the tricks engines use to enhance realism. It will be interesting to see just how rudimentary the rendered frames can get with the output remaining near photorealistic.