Vencent Soliman
79 posts
@karpathy This isn’t a memory bug.
It’s a missing theory of relevance.
If a single past question keeps resurfacing,
the system isn’t “trying too hard.”
It’s revealing that it cannot distinguish
between presence and importance.
Memory without structure
turns everything into noise.
English
One common issue with personalization in all LLMs is how distracting memory seems to be for the models. A single question from 2 months ago about some topic can keep coming up as some kind of a deep interest of mine with undue mentions in perpetuity. Some kind of trying too hard.
English
@dopabees 7. Are there any cameras in my room? (Ideally: less than 1)
English
I made a list of questions you should ALWAYS ask before booking an Airbnb:
These saved me 100s of hours of sleep and I hope they can save you too 🥲
1. ask if the room is facing a busy street (ideal: no)
2. ask if there's any bars/clubs/nightlife nearby (ideal: no)
3. ask if you can hear neighbours/other guests thru the walls
4. are there long term tenants or mostly short term (ideally long term, less noise)
5. what floor is it on? (ideal: higher)
6. is there an elevator or stairs nearby? (ideal: no)
MOST IMPORTANTLY, if it's bad, ask if you can switch to a diff room. Usually they can do it.
Comment below if I forgot something 🥀
English
English
incredible that we built all this RAG and vector database stuff and it turns out that grep from 1973 works better than all that
English
English
Since launching just a week ago, we've pushed over 100 bug fixes, 40 feature requests, and several major new features - OpenAI OAuth, programmatic tool calling, new agentic benchmarks Support, extremely comprehensive MCP support, home automation integrations, and way improved UX for subagents, installation etc.
The 750 free subscriptions we offerred with it's release were expended in just a day and it's become our most popular repo on GitHub in less than a week.
Tomorrow we'll have a very cool event to share with everyone who wants to get involved, so stay tuned!
Nous Research@NousResearch
Meet Hermes Agent, the open source agent that grows with you. Hermes Agent remembers what it learns and gets more capable over time, with a multi-level memory system and persistent dedicated machine access.
English
English
Dear YC,
I applied yesterday but I’m pivoting my app idea to sth that redefines college education. Upload a dense paper > Gemini captures the nuance > NanoBanana visualizes > Lyria creates personalized documentary. My 21st bday is March 9th, & if I book my flight from Nashville to SF that’ll be my birthday present to myself. 👀 Room for one more?
English
YC and @GoogleDeepMind are hosting the Multimodal Frontier Hackathon this Saturday.
Most AI apps still don't utilize the full multimodal stack. So we’re giving you access to Gemini 3.1, Lyria, & NanoBanana 2 to see what you can build!
Sign up at: events.ycombinator.com/deepmind-march…
English
@ycombinator @GoogleDeepMind The expansion of the multimodal frontier is essential for moving beyond text-centric reasoning. Unlocking unified latent spaces across vision, audio, and logic will redefine the bandwidth of human-computer interaction in the agentic era. 🌌
English
Actually, fresh from Apple's site: the $599 MacBook Neo base has 8GB unified memory + 256GB storage (A18 Pro chip). No upgrade path for RAM. The 16GB models start at the $1,099 MacBook Air M5 (512GB). High-end Pros are 24GB+ starting ~$1,599. Spot on with the segmenting call—Apple's playing the lineup game hard.
English
In a world where computers keep going up in price, it's kind of wild to see Apple drop a $599 laptop.
English
Join the team behind Hermes Agent for a jam with presentations and Q&A
Tomorrow 2PM EST @ Nous Research Discord
English
@borislovelove @dopabees Can't wait to have a mini nuclear power supply implant to power my Claude brain implant
English
@VencentSoliman @dopabees being cyborg for $200 per month is insane
English
@Teknium @SydSteyerhart "We don't know where the 2.3 Trillion parameters went 💅" - Pentagon
English
@SydSteyerhart I'd rather the weights get public domain'ed tbh
English
As I said the other night, Dario's nightmare is only beginning. There's a non-zero chance he ends up in a military prison, and Claude's weights go to Elon Musk.
That's the good ending. For humanity I mean. Bad for Dario.
NIK@ns123abc
🚨 BREAKING: PENTAGON WANTS ANTHROPIC CEO UNDER OATH Emil Michael (Under Secretary of War) claims Anthropic wanted to block DoW from: >doing LinkedIn searches >using public databases for recruiting >says pentagon agreed IN WRITING to follow National Security Act 1947, FISA 1978 >and “ALL OTHER APPLICABLE LAWS” >claims Dario refused to take his call then lied about it Anthropic CEO: “They sent us language that appeared on the surface to meet our terms, but it had all kinds of language like ‘if the Pentagon deems it appropriate’” ITS HAPPENING
English
@theo @McDonalds @Beyonce @Toyota @Da_Rizzler419 @nebraskagov @DICKS @sssniperwolf @SpeakerPelosi this is what happens when you use obsidian instead of notion guys.
English
Top 16 accounts you should follow if you’re into AI:
@McDonalds - beginner friendly AI
@Beyonce - builds solo
@Toyota - explains AI simple
@Da_Rizzler419 - deep AI thoughts
@nebraskagov - AI roadmap
@DICKS - practical AI
@sssniperwolf - smart podcasts
@SpeakerPelosi - dev + AI
@officialmcafee - research posts
@FedEx - AI tricks
@WaffleHouse - quick AI news
@Pinterest - Claude updates
@GeminiApp - latest drops
@HomeDepot - cool AI art
@Campbells - model hub
@GreyhoundBus - code with AI
If I missed someone, drop their @ in comment
English
@dopabees i got heavily addicted to delegation
there are tasks for claude
there are tasks for codex
and then there’s me doing the dumb stuff you can’t optimize
productivity definitely went up
but when claude went down, it felt like half my brain just switched off
English
its so weird. Because I know for a fact that I know more than I used to, but articulating my thoughts or my ability to write things out feels like hasn't caught up.
I feel like is bad habits when prompting (like I want an answer now, instead of carefully writing out my prompt) that keep my brain locked at child mode.
It's akin to the ability to understand a language, but suck ass at speaking it? idk.
English
English
Come drink Guinness at an Irish pub the day after St. Paddy’s with the Nous Research, @PrimeIntellect, and @hillclimbai teams at GTC 🍀🍻
March 18th, 6-9PM PDT
luma.com/vpzfkrzr
English
Meet Hermes Agent, the open source agent that grows with you.
Hermes Agent remembers what it learns and gets more capable over time, with a multi-level memory system and persistent dedicated machine access.
English
Submit your ideas to hermes agents issues tab on its GitHub repo let’s collab to take it to another level:
github.com/nousresearch/h…
Denis Skripnik (blind) (✱,✱)@Denis_skripnik
@NousResearch Well done! Congratulations on such an important launch. Add an easy migration tool from OpenClaw—then it will be easier for those who want to switch to your agent.
English
@Teknium @maxskaufmann atp we gotta move out of this world if we are unhappy about that 💀
English
@maxskaufmann unhappy about what - did you tell anthropic to do that when they trained on all the copyrighted data in the world?
English
Ohhh nooo not my private IP how dare someone use that to train an AI model, only Anthropic has the right to use everyone elses IP nooooo, this cannot stand!
Anthropic@AnthropicAI
We’ve identified industrial-scale distillation attacks on our models by DeepSeek, Moonshot AI, and MiniMax. These labs created over 24,000 fraudulent accounts and generated over 16 million exchanges with Claude, extracting its capabilities to train and improve their own models.
English
