Alex

[Alex @mur3ai]

I closed this line as a negative result, with one useful finding: the forgetting has two causes. During training, the target's excitation rises until it no longer discriminates; at rest, the learned weights decay by about 30%. They likely need separate fixes.

[Alex @mur3ai]

Who actually copy-pastes in 2026? You know agents exist, right?

[Sam Ongaro @samongaro_]

Developer honesty test:

Do you understand how your code works, or are you just copy-pasting what AI generates?

[Alex @mur3ai]

No, It does not make me believe I'm an expert. But I do believe that AI can handle all the tasks an expert can and even better at it than most human self proclaimed experts.

[Simon Maechling @simonmaechling]

AI is just a way to make non-experts believe they are experts on topics they simply don't understand.

[Alex @mur3ai]

Yeah, my Claude does. And if if it does not there are Astra and Argon to help him. Actually my prod became a lot MORE stable with vibecoding than before when it was written by human programmers.

[Juan Pablo @master_jpma]

Vibe coding is fun until production breaks.

Do you understand the code well enough to fix it?

[Alex @mur3ai]

This one surprised me: with learning left on, recall strength fell by 79%, and 40 further rehearsals of the same pairs erased all 20. The network forgets because it keeps learning. My best candidate fix passed on at most 5 of 20 fresh networks.

[Alex @mur3ai]

The rule that worked: strengthen a synapse when both neurons are above their mean, weaken it three times as strongly when they diverge. Recall on 20/20 fresh networks, 0/20 in controls, stable over 10k frozen steps. Replacing learned pairs still fails.

[Alex @mur3ai]

Adding temporal traces, competition, adaptive thresholds and consolidation brought recall to 3 of 10 fresh networks; I required 8. The main problem was specificity: in about a quarter of probes, unpaired cues also activated the target.

[Alex @mur3ai]

My first memory test: pair cue A with target B, then present A alone and check whether B is recalled. Across 120 conditions, 0 of 10 networks recalled. The weights changed, but no usable memory formed. Not the start I had hoped for.

[Alex @mur3ai]

I'm building an agent that learns from its own experience, with no LLM inside. The substrate is a sparse recurrent network of 10,000 neurons on a laptop GPU; each synapse learns only from its own two neurons, using a bounded version of Oja's (1982) rule.

[Alex @mur3ai]

best claude skills come from pure routine. you just go "automate this thing we've been doing by hand for 4 days so you can do it yourself" and it does. kinda boring now tho

[Alex @mur3ai]

working for a ceo with cyberpsychosis is rough. treats you like an agent and wants everything shipped in 10 mins. ok sure lets yeet all the processes and push to prod with zero QA and zero monitoring

[Alex @mur3ai]

gave claude a skill to manage free muse spark 1.3 in opencode and it turned into a seagull manager. keeps popping in like "any update?" "eta??"