Nils Höglund

Tankar om gymnasielagendomarna @Fossey/tankar-om-gymnasielagendomarna-bab180a17d6a" target="_blank" rel="nofollow noopener">medium.com/@Fossey/tankar…

@lejooon @Investeraren @montrose_io Ja. Men knte lika enkelt provisionera idag till företag med ex hundratals utvecklare i olika editormiljöer som MCP?

Samma sak kan du åstadkomma med instruktioner och API:er. Håller inte med om att det är enkelt då det kräver mycket kontext. Det är ingen magi och i praktiken kan du emulera en MCP mha en API om du ville. MCP tillkom för att standardisera vissa saker när LLMs var sämre på att följa instruktioner.

BREAKING NEWS‼️🔥 Montrose är först ut i Europa med en MCP-server för aktier och fonder. Det gör det möjligt att koppla din depå till valfritt AI-gränssnitt och dessutom utveckla egna verktyg 🥳🙌 ✅ Prata med din portfölj ✅ Analysera dina innehav ✅ Skapa bevakningar ✅ Initiera köp ✅ Utveckla egna verktyg (dashboards m.m.) ✅ ... och mycket mycket mycket MYCKET mer. Fantasin sätter gränserna!! Detta markerar ett tydligt skifte för svenska investerare. Vi fortsätter att driva innovationen i branschen. Montrose - av investerare, för investerare! Först ut i Europa avser banker, nätmäklare och värdepappersbolag och aktier/fonder. Alltså första aktören som tillåter sina kunder att dra nytta av kraften med AI. #PrataPengar #Finanstwitter montrose.io/mcp

@lejooon @Investeraren @montrose_io Tyvärr kommer kontext-”problemet” med MCP gissningsvis snart ”lösas”. Vi ser ju redan anpassningar av responses api för att börja ”lösa” detta: developers.openai.com/api/docs/guide… Då kommer man inte ha samma controllability över hur tasks löses lika enkelt som idag genom MCP-contextbagaget

@lejooon @Investeraren @montrose_io Står fast av viss fördel - eftersom det är ngt du (som skapare av MCP-servern) kontrollerar och kan guida LLM med till viss del i hur uppgifter ska lösas. Svårt att göra annars på så idag kompatibelt sätt med så enkel provisionering

@lejooon @Investeraren @montrose_io Men just därför ska tredje parter hålla sig borta och i stället erbjuda vanliga API:er (som MCP och annat kan koppla upp sig emot)..

@lejooon @Investeraren @montrose_io Att MCP slukar kontext normalt är en *fördel*. Det ger enkel controllability över LLM och inriktningar för den som designar MCP-servern. Perfekt för interna verktyg, uppkoppling mot API och väldigt enterprise-vänligt med enkel provisionering och hyfsad kompatibilitet.

For the record, the president of the United States is now simultaneously claiming that he has won the war, is currently winning the war, needs help to win the war, and needs no help to win the war. All to destroy the nuclear program he claims to have already destroyed last year.

@BuschEbba Suck. I rådande säkerhetspolitiska läge bör Sverige undvika att ansluta sig till USA-drivet blocktänkande. I stället bör vi inta en avvaktande hållning för att bevara största möjliga framtida handlingsfrihet.

Good news! 🇸🇪🇺🇸

@BetterCallMedhi How did he get the mRNA vaccine manufactured? Who helped him with that?

I genuinely think this might be the most important story I’ve read this year and I need to talk about it a guy in australia just designed a custom mRNA cancer vaccine for his dying dog using chatGPT and alphafold, he has 0 background in biology and it worked, the tumor shrunk by half, the genomics researchers are absolutely stunned & I genuinely think this story is way bigger than people realize here’s what he actually did, he paid 3000 bucks to get the tumor DNA sequenced, fed the data to chatGPT to identify mutations of interest then used alphafold to predictt the 3D structure of the mutated proteins & find therapeutic targets, then he designed a custom mRNA vaccine targeting the specific neoantigens of his dog’s tumor, all of this from his laptop & the genomics professor who received the sequencing request initially thought it was a joke few months later this same professor is looking at the results saying if we can do this for a dog why are we we rolling this out to all humanswith cancer and this is where I need you to understand what alphafold actually represents because I’m convinced most people have heard the name without grasping what’s hapening underneath: for decades figuring out the 3D structure of a single protein required months sometimes years of X-ray crystallography /cryoelectron microscopy, entire labs dedicated to one molecule, alphafold2 solved this by predicting the structure of virtually every known protein thats ovr 200 million structures which earned it the Nobel prize in chemistry in 2024 but here’s the thing, alphafold 3 released in 2024 went even further where alphafold 2 predicted the structure of an isolated protein alphafold 3 predicts interactions between proteins DNA RNA small molecules & ligands in a unified system basically it models how a drug molecule will bind to a protein target with 50% better accuracy than the best existing tools & it does it in hours instead of years and thats exactly what this guy exploited for his dog, he used alphafold to see the 3D shape of of the mutated tumor proteins & figure out how an mRNA vaccine could teach the immune system to recognize & destroy them specifically and look what fascinates me personally is what this signals for whats coming next isomorphic Labs the deepmind subsidiary dedicated to drug discovery already signed multibillion dollar partnerships with Eli Lilly & Novartis and the first drugs entirely designed by AI through alphafold3 are expected to enter human clinical trials by end of 2026 we’re talking oncology & immunology candidates that were designed through rational design meaning the AI literally drew the molecule to fit perfectly onto the target instead of screening millions of random compounds like we’ve been doing for 50y by the way the movement is accelerating way faster than people think, deepmind open sourced alphafold 3 in late 2024 the scientific community immediately built on top of it, models like OpenFold3 backed by amazon & Novo Nordisk, startups like recursion developing specialized versions… I’m telling you we’re entering the era of the autonomous lab where AI designs a molecule robots synthesize it & high-throughput platforms test it with 0 human intervention I believe the next frontier is temporal modeling, today alphafold predicts the static shape of a molecule tomorrow we’ll predict how it moves & vibrates over time inside a living cell & after that come patient digital twins simulations that predict how your specific genetic variations will affect your response to a given drug, truly personalized medicine at the atomic level traditionally it takes 15 years & roughly billion dollars to bring a drug from discovery to market, AI is compressing that cycle at a pace that should terrify every incumbent & what this australian guy just proved is that the entire pipeline tumor sequencing target identification structure prediction custom vaccine design can be executed by 1 person with a laptop for a few thousand $$

Ligger något i detta. Många kommer att gå i denna fälla och mycket kommer att bli felallokerat i ekonomin kommande år för att state of the art gällande detta vid olika tidpunkter inte är allmänt förstått.

@KingBootoshi Just make sure your invariants capture meaningful and non-trivial properties and that your specification is at the right level of abstraction for the properties you need to verify…

HOLY FUK I JUST LEARNED ABOUT TLA+ AND IT'S SO GOOD FOR AGENTIC CODING ur telling ME that i can mathematically fact check every possible scenario of my design STATE to prevent bugs and crashes AND IF IT FINDS SOMETHING THE AGENTS GET INSTANT FEEDBACK AND LOOP FIXING IT TILL IT ALL POSSIBLE BUGS IN THE DESIGN ARE PATCHED LOL THIS IS OP

@ankrgyl The main benifit of MCP is easy deployment/updates (one URL behind oauth - great for enterprises), together with wide compatibility in the AI tools ecosystem (different AI-assistant vendors/OS:es/etc). ”Just works” if you use the basic features.

people who feel strongly on mcp vs. cli i am doing some research here. are there any particular services that have an mcp and cli (both) and you feel like one is way better than the other?

@Ride_fixer Krya på dig!

Om ni undrar över minskad närvaro beror det på att jag har fått influensan från helvetet.

@lanreadelowo Yeah. But you have to learn what you want done. Easier said than done

“Learn Claude skills” My brother in Christ, it’s just a text file where you have to explain what you want done

@fatjudge (Gällande hovrätt: förutom vid ovanliga tillfällen då presidenten är med och dömer.)

Vem som är chef för en viss domstol har självklart noll betydelse för hur domstolen dömer i ett visst fall.

@sama ChatGPT 5.4 Pro thinks for 88 minutes and generated a link to a file. Great, except for file link expired by the time the link is delivered. Re-occurring problem that has persisted forever. Please fix this basic think in your premium product?

Det speglar verkligheten på vissa sätt. Intelligens är inte tillräckligt för att lyckas i livet, omvänt är det inte heller nödvändigt. Handlingskraft, driv och initiativförmåga betyder minst lika mycket.

@fatjudge Vi får se om riksdagen (i enlighet med NJA 2018 s. 743 p 39) använder det försteg de har att uttolka regeringsformen och redovisar sin bedömning av Lagrådets yttrande.

Mycket från lagrådet just nu. Deras hemsidan är att rekommendera om man vill hålla koll på det. lagradet.se/yttranden/ Här ett tydligt besked om förslaget om förlängd tidsgräns för häktning av omyndiga. Som även skulle kunna gälla 13 och 14 åringar.

Det är lite som att skryta att deras utvecklare inte har skrivit en enda rad assembler eller maskinkod sedan december - det har deras kompilator gjort åt dem.

Leadership: "AI is so good it writes almost all our code" Status dashboard:

@protosphinx The models can ”learn”. It is called the context window. The key is to extend that with still perfect recall.

AGI is not coming. We are nowhere near AGI. What we have today is inference, not learning. Models get trained once on huge fixed datasets, then frozen. You ask questions, they remix patterns they already saw. Nothing updates. Nothing sticks. Talking to the model does not make it smarter. It does not learn from you. Ever. Learning is still slow, expensive - and offline. Look at self driving. You drive around a pothole, make a U turn, and come back. The car’s AI does not learn that you just solved that exact problem. It reacts the same way every time using sensors and rules. Do this 20 times a day and it still has zero memory that the pothole exists. It just re sees it. That is why edge cases never die. There is no local learning. No accumulation. No 'oh yeah, I’ve seen this before' LLMs work the same way. Tell it your name and it does not remember. The only reason it looks like memory is because scaffolding keeps shoving your name back into the prompt every time and sanitizing the output. The model itself has no idea who you are and cannot learn from interaction. It is structurally incapable. And the scaffolding is the worst part. It is pure duct tape. Just prompts on prompts on prompts around a frozen model. When something breaks, nobody fixes learning. They add another layer. Another rule. Another retry. Another evaluator model judging the first model. So you end up with systems that are insanely complex but mentally shallow. Debugging is hell because behavior comes from hack interactions, not a learnable core. Tiny prompt tweaks cause wild behavior shifts. Latency goes up. Costs go up. Reliability goes down. None of this compounds into intelligence. It just hides the cracks. Until we have real persistent learning and real memory inside the system, there is no AGI. LLMs are not built for this. You cannot prompt your way out of it. You need a totally different architecture. Yann LeCun is right. And even then, what architecture can actually learn online, store memory, and stay stable on today’s hardware? Best case, maybe 5-10 yrs. Right now it is all inference. It looks magical, but the emperor has no clothes. A lot of people see it. Almost nobody says it out loud.