Advice for AI makers

A friend of mine is about to launch him­self heav­ily into the realm of AI pro­gram­ming. The de­tails of his ap­proach aren’t im­por­tant; prob­a­bil­ities dic­tate that he is un­likely to score a ma­jor suc­cess. He’s asked me for ad­vice, how­ever, on how to de­sign a safe(r) AI. I’ve been point­ing him in the right di­rec­tions and send­ing him links to use­ful posts on this blog and the SIAI.

Do peo­ple here have any recom­men­da­tions they’d like me to pass on? Hope­fully, these may form the ba­sis of a con­densed ‘warn­ing pack’ for other AI mak­ers.

Ad­den­dum: Ad­vice along the lines of “don’t do it” is vi­tal and good, but un­likely to be fol­lowed. Cod­ing will nearly cer­tainly hap­pen; is there any way of mak­ing it less geno­ci­dally risky?