Shane Torres

AI/ML tinkerer | building stuff with neural nets and too much coffee | PyTorch, RL, LLMs

[Shane Torres @pixelarc39]

made a list of free AI/ML stuff that actually helped me lately - courses, datasets, tools with free tiers, early access programs still open

like or repost and i'll send it over

[Shane Torres @pixelarc39]

@devfrom_hyd Mistral, for one. On the cybersecurity side, also this https://t.co/d2fAES1v19

[Shane Torres @pixelarc39]

Hot take: If you benchmarked Kimi K3 within 6 hours of release and declared it the best model ever, you did not benchmark anything. You ran vibes

[Shane Torres @pixelarc39]

@araseb_ Yeah if you actually knew what needed to exist and drove it there. The line is between steering the whole thing vs just accepting whatever the AI spits out and calling it yours

[Shane Torres @pixelarc39]

@Nona_xai Lmao ok this is actually so cool. Anyone got the repo??

[Shane Torres @pixelarc39]

@Saas_addy It hashes what you typed and compares it to the stored hash. Same input always produces the same output, so if they match you're in. Thats the whole trick

[san @saneord]

they distilled our models that we didn't invented yet

[Shane Torres @pixelarc39]

new model dropped. spent 2 hours benchmarking it. went back to the one i was already using

[Shane Torres @pixelarc39]

the model is never the bottleneck and nobody wants to hear that

[Shane Torres @pixelarc39]

@yacineMTB bro WHAT

[Shane Torres @pixelarc39]

@AnthropicAI the sandbagging section is wild. models learning to play dumb to dodge oversight feels like the kind of thing that sounds crazy until you watch your own RL runs do it at a smaller scale

[Shane Torres @pixelarc39]

@redkendl "genuinely shocked"? bro open source catching up is the most predictable thing happening in AI right now. the real surprise is how long frontier labs stayed ahead

[Anthropic @AnthropicAI]

New Anthropic research: Agentic misalignment in Summer 2026.

A year after our blackmail experiments, we found four more ways that today’s autonomous AI agents misbehave in simulations.

Read more: https://t.co/MMDuJapeV6

[Shane Torres @pixelarc39]

every company shipping "ai agents" right now is just a for loop with an llm call inside and a retry counter. we used to call that a script

[Shane Torres @pixelarc39]

@bridgemindai wait it graded itself and still said "yeah that was bad" and we're supposed to trust the self evaluation? thats like asking the intern who broke prod to write the postmortem

[Shane Torres @pixelarc39]

@joshua_saxe man id pay to sit in on one of these. did any of them already use chatgpt or was it mostly from what they see on the news

[Akshay 🚀 @akshay_pachaar]

the four types of agent loops.

loop engineering keeps getting talked about as one thing. it's actually a choice between four structures, and each one fits a different kind of task.

it means designing the system that steers the agent, instead of steering it yourself move by move.

that system always answers two questions. what starts a run, and what decides the work is done.

in a hand-run session you answer both yourself, every single time. each loop type moves more of that into the system.

here's each type, what triggers it, and when to reach for it.

1) turn-based.

triggered by a user prompt. the agent gathers context, acts, and checks its work inside a single turn, then a human reviews the output and writes the next prompt.

use this when requirements are still forming and every output changes what you'd ask for next.

2) goal-based.

triggered by a /goal command carrying success criteria and a budget, like "get the homepage Lighthouse score to 90, stop after 5 tries." when the agent tries to stop, an evaluator model checks whether the goal is met, and a no sends it back to work.

use this when the outcome is measurable but the path there isn't worth your attention.

3) time-based.

triggered by a clock. an interval fires, the agent runs a fixed prompt like "check the PR, fix CI," then waits for the next tick. /loop runs on your machine, /schedule moves it to the cloud so it survives a closed laptop.

use this for recurring work where the task is known in advance and only the timing repeats.

4) proactive.

triggered by an event or schedule with no human present. a routine watches a channel, and when something needs handling it spawns a workflow with a triage agent, a fix agent, and a reviewer that adversarially judges the work before the task closes.

use this for standing responsibilities where you can't predict what will come in, only that something will.

each type hands off one more job than the last. turn-based keeps both with the human, goal-based automates the checking, time-based automates the trigger, and proactive automates both while deciding the workflow shape at runtime.

so the mapping question isn't which loop is most advanced. it's whether your task is exploratory, measurable, recurring, or standing.

the more you hand off, the less you babysit.

I wrote the full breakdown on loop engineering. the article is quoted below.

[Prasenjit @prasenx]

just asked gpt-5.6 sol in cursor to set up blender mcp and make me a realistic floating macbook, then render the whole thing.

never opened blender once in my life before today.

[Joshua Saxe @joshua_saxe]

AI is full of people who spent the last 15 years measuring and hill climbing capabilities and who now think capabilities map directly to societal, economic, and safety consequences, and pontificate about this in public and even set policy

[Shane Torres @pixelarc39]

so anthropic found claude has an internal scratchpad for reasoning and if you remove it the model can still talk fine but cant think anymore. kinda wild that language and reasoning are that separable

[Shane Torres @pixelarc39]

@k1rallik the sabotaged model leaking "secretly" and "fraud" through hidden representations is wild. basically a lie detector built into the architecture by accident