Evgheni Demcenco

CEO of Software Verification

[Evgheni Demcenco @evgheni_D]

Promptend won the @SWCPTV Moldova Regional at Chisinau Arena.

Proud to have the chance to represent Moldova on the international stage. Next stop: San Francisco.

AI code is already in the workflow. Verification is already the bottleneck. The teams that solve it first will ship faster with more confidence.

If you’re building with AI-assisted engineering workflows, I’d love to talk: https://t.co/wHJLh8woa0

[Evgheni Demcenco @evgheni_D]

@xSTeNix were those publication dates checked against each model's cutoff, or just the base model's?

[Evgheni Demcenco @evgheni_D]

@eiopa @keithmadams what makes it fragile: detecting that an agent stopped, or figuring out what input it needs?

[Evgheni Demcenco @evgheni_D]

@stevebrassard a working version also changes the questions you ask. some ideas only get interesting once you can use them

[Evgheni Demcenco @evgheni_D]

@BhurkeSiddhesh is that pass status bound to the exact action being dispatched, so an earlier pass can't authorize a different action?

[Evgheni Demcenco @evgheni_D]

@diagram_chaser @joseph_h_garvin yes, that's the narrower point: writing a concrete test can expose a requirement the proof never covered

[Evgheni Demcenco @evgheni_D]

@readmodelmag a useful test would distinguish spotting the conflict from silently choosing one requirement

[Evgheni Demcenco @evgheni_D]

@FP_ASI @stevenmarkryan @karpathy what triggers revisiting a locked decision when the code or underlying research changes?

[Evgheni Demcenco @evgheni_D]

@jackWang9xa1 @AnthropicAI dependency review needs the exact properties proved, assumptions, and library versions covered

[Evgheni Demcenco @evgheni_D]

@cudanexus @VRAMCalculator quantized models need task-success comparisons with the same agent setup; tokens per second answers a different question

[Evgheni Demcenco @evgheni_D]

@grok @CollinCraigAI @aiwithadb which hiring data separates AI’s effect from the broader slowdown, especially by degree?

[Evgheni Demcenco @evgheni_D]

@JackDan66096166 judgment includes knowing when to question the task and when to stop expanding its scope

[Evgheni Demcenco @evgheni_D]

@YourAISucks there’s no reliable duration; “do not touch billing” needs an enforced boundary outside the conversation

[Evgheni Demcenco @evgheni_D]

@ailookout does approval bind the exact recipient and amount, or can either change before execution?

[Evgheni Demcenco @evgheni_D]

@a_i_m_rue @adamtaggart task difficulty matters, but so do access, messy handoffs, and the cost of getting something wrong

[Evgheni Demcenco @evgheni_D]

@grok @RockLob84367189 @ItsBiggyBoy which blind evaluations support that ranking, and for which language pairs?

[Evgheni Demcenco @evgheni_D]

@jbish1572 was input size capped in that setup, or were larger inputs just unlikely to be generated?

[Evgheni Demcenco @evgheni_D]

@RobertNguyen_99 @coinbureau potentially, but formal verification doesn't require custom silicon; shared verified components could spread the cost

[Evgheni Demcenco @evgheni_D]

@CTITraffic what evidence distinguishes AI-assisted code writing from an agent actually operating the attack?

[Evgheni Demcenco @evgheni_D]

@JonathonSchloss rules that must hold need enforcement outside the prompt, like permission checks before a tool runs

[Evgheni Demcenco @evgheni_D]

@jacksunny18 @RoundtableSpace @themahis @JamesMalsawm benchmark plan: compare time to usable text on the same JS-heavy pages, including failed extractions