Kimiyyar kwamfuta
Harshen da bai shafi yadda kuke tunani game da shirye-shirye ba, bai cancanci sani ba.
— (1982)
Gogaggun masu haɓaka software sun ji kamar AI ta sa su sauri, alhali ma’auni ya nuna sun yi jinkiri — wannan ne ainihin sakamakon, ba hukunci kan AI coding gaba ɗaya ba
A randomized controlled trial na METR ya sa gogaggun masu haɓaka open-source 16 su yi ayyuka na gaske 246 a codebases da suka sani sosai, tare da kayan AI na farkon 2025 a rabin ayyukan da aka rarraba bazuwar. Sun yi tsammanin AI za ta rage lokaci da kusan 24%; maimakon haka lokacin kammalawa ya ƙaru da 19%, kuma bayan gwajin har yanzu sun yi imanin AI ta hanzarta su da kusan 20%. Wannan tazarar fahimta ita ce babbar zuciyar binciken, amma mutane 16 ne, yanayi ɗaya, da hoton kayan aikin farkon 2025 — ba hukunci kan duk AI coding ba.
Cryptographic proof zai iya ɓoye sirrinsa ta sa missing simulator ya zama mai wahalar prove
Zero-knowledge proofs suna ba mutum damar prove statement ba tare da bayyana witness ba. Classical theory tana cewa ba za ka iya samun wannan, a full sense, da message guda, babu trusted setup kuma perfect soundness ba. Paper na Rahul Ilango bai sa wannan impossibility ya ɓace ba. Yana canza target: maimakon buƙatar simulator ya wanzu da gaske, yana tambayar cewa chosen proof system ba zai iya efficiently prove cewa simulator ba ya wanzuwa ba. A ƙarƙashin manyan assumptions na proof complexity da cryptography, wannan ya isa a dawo da falsifiable, game-based consequences na zero-knowledge property by property. Point ɗin ba plug-in internet primitive ba ne. Proof-theoretic hanya ce ta juya mathematical unprovability zuwa cryptographic cover.
Medical AI na iya zama private on average amma har yanzu ya expose particular patients — especially underrepresented
Medical-AI model da ake kira “privacy-preserving” yawanci yana dogara da average number guda. Wannan study yana cewa wrong test ne. Ta studying membership inference attacks — waɗanda ke reveal whether specific person's record yana training data, don haka zai iya betray disease status — authors sun measure risk per patient rather than aggregate, across seven medical datasets (imaging, ECG, electronic records) da many models each. Pattern: models da look safe on average na iya still let attacker identify specific individuals almost perfectly (attack AUC ≥ 0.95); exposed systematically from underrepresented groups; kuma gets worse as models grow. Authors ba su ce abandon medical AI ba — suna ce measure privacy per patient, control model access, use differential privacy. Core: “private on average” ba privacy guarantee ba ne.
Koyar da drawing AI ta kalli page yayin da take zana
Language models da ke generate vector graphics sun saba yin shi blind — suna rubuta drawing commands ba tare da ganin result ba. Sabuwar method tana sa model ta kalli canvas stroke by stroke, amma twist ɗin shi ne simply ba ta eyes yana sa quality worse: dole a retrain ta yi amfani da visual feedback. Bayan training, model tana match ko edge out rivals da aka train da har 20× more data — real, careful result a benchmark ɗaya, ba revolution ba.
AI agent ya gudanar da entire patient cases da kansa — a simulator, on past records
MIRA new kind of medical AI ne: maimakon answering single question, yana work entire case inside simulated hospital record — taking history, ordering da reading tests, reaching diagnosis, writing orders. On 574 retrospective cases from public database across eight pre-selected diagnoses, authors report it outperformed physicians on diagnostic accuracy and made largely guideline-concordant, medication-safe decisions. Every qualifier matters: sandbox on past records, text only; much edge on clear-cut test conditions; ordered about twice as many blood tests as doctors; several outcomes scored against original chart. Genuine advance agent that acts across whole workflow — not proof machine diagnoses better than doctor. Authors say generalization, safety and governance still need prospective real-world studies.
Shin manyan robot “foundation models” suna aiki fiye da gaske? Amsa mai hankali — eh, kaɗan, kuma yawancin studies ba za su iya gane bambancin ba
Toyota Research Institute ta horar da “large behavior models” — robot policies da aka pretrain a kan ~awa 1,700 na diverse manipulation data — sannan ta gwada su da from-scratch single-task policies da unusual rigour: blind, randomized, large-sample trials (~1,800 real-world, 47,000+ simulation) tare da real statistics. Bayan per-task finetuning, big models sun fi kyau a matsakaici, sun buƙaci kusan 3–5× ƙarancin task-specific data, kuma sun fi robust lokacin da conditions suka canza; performance ya ƙaru smoothly da karin pretraining data. Amma ba tare da finetuning ba ba su consistently beat single-task models ba, effects da dama sun yi ƙanƙanta har sai da manyan samples suka bayyana su, kuma data-normalisation choice na yau da kullum ya fi architecture tasiri. Wannan measured support ne ga robot-foundation-model direction — ba general-purpose robot ba, ba zero-shot generalist ba, ba “emergent leap” ba — tare da gargadi cewa wani ɓangare mai yawa na robotics na iya kasancewa yana auna noise.
Me ya sa model na harshe suke yin hallucination — da kuma me ya sa yadda muke ba su maki yake ci gaba da haifar da shi
Ƙaryar da ake faɗa da amincewa da muke kira “hallucination” ba wata matsala ta asiri ba ce: wasu sakamakon ƙididdiga na horo ne, kuma suna ci gaba domin manyan gwaje-gwaje suna ba zato mai ƙarfin hali lada fiye da “ban sani ba” na gaskiya. Wani gwajin misali a kan manyan model huɗu ya nuna cewa faɗin ƙa'idojin maki a cikin umarni (“ƙa'idoji a fili”) yana juya wannan lada.