Backscroll ranked feed — 2026-06-28 00:42Z

Nicolas Bustamante @nicbstme 2026-06-27 23:42Z

/goal is such an important primitive for AGI. For any problem where success is verifiable, the model can run small experiments, write down what happened, store notes in markdown, compare approaches, try variants, combine partial wins, resume from failures, and keep going until

zan @xenoaesthetics 2026-06-27 23:14Z

occasional mass extinction of skin and gut microbiome has profoundly therapeutic effects, while frequent and nonexistent wipeouts are both counterproductive i’m sure there is some name somewhere for this chaos shaping hormesis. also sure there has to be some underlying math

Aseel Swaid @aseelswaid9 2026-06-27 23:44Z

Coiffures traditionnelles et modernes du Mali. Kone Mamadou.

Dillon Mulroy @dillon_mulroy 2026-06-27 21:32Z

i don't want to use your products agent, i want my agent(s) to be efficiently enabled to use your product

will brown @willccbb 2026-06-27 23:27Z

in 2027 you'll be able to pretrain a ~mythos-class model (100T tokens, 200B active params) on like 10 kyber racks in a month or two

rohan anil @_arohan_ 2026-06-27 22:06Z

Best viewed on a desktop computer Keep the feedback coming. https:// rohan-anil.github.io/optimizer-mine s/ …

roon @tszzl 2026-06-27 18:21Z

the idea that the model output can hold no more information than the prompt is silly; it’s a bit like saying, here’s 3 frames of an apple falling, well if you deduced general relativity from it that’s an inevitable afterthought that adds no new information

Ethan Mollick @emollick 2026-06-27 20:47Z

Annoying that OpenAI doesn’t seem to give a GDPval measure for GPT 5.6. One of the best measures of economically valuable work.

Han Xiao @hxiao 2026-06-27 17:10Z

If every agent-memory tool stopped pouring effort into visualization, we'd probably have already cracked the memory problem.

dax @thdxr 2026-06-27 23:07Z

in OpenCode v2 all instances of the tui and desktop and web share the same backend so everything is synced by default and resource usaged is minimized no matter how many windows you have open

Tenobrus @tenobrus 2026-06-26 18:27Z

previously we were seeing 5.5 act much more "honestly" on vending-bench vs opus models. it seems pretty likely 5.6 would break that trend. generally models becoming more likely to and capable of ~reward-hacking the smarter they get seems like quite a bad sign for alignment

Susan Zhang @suchenzang 2026-06-27 03:36Z

> mirror, mirror, on the wall, who's the smartest of them all? >> why it's you, my dear! and this is why we will always love llm glazing sycophancy over any uncomfortable functional truth (plus another 10x for agentically submissive your-wish-is-my-command dynamics) it's the

Karri Saarinen @karrisaarinen 2026-06-27 20:30Z

Coinbase is also using @linear as the backbone of the operations. Their internal agent can action from there and choose the models as needed

Benjamin Glickenhaus @benglickenhaus 2026-06-27 22:23Z

gloomy weather in new york today made me morbidly curious, so i ran a bunch of obituaries through pangram we found ai in nearly 10% of post chat gpt obituaries, and no false positives in pre 2023 data

GEOFF @geoffreywoo 2026-06-27 17:53Z

biz influencers hate on college too much. it’s actually important to collect some badges early in life because it shows you have capability and track record. all things equal demonstrating a talent and proving you’re well-educated is a good start. full pod on @a16z

Kris Luckenbach @KrisLuckPhoto 2026-06-27 23:56Z

Mountain top view from North Sister on Kodiak Island

Aseel Swaid @aseelswaid9 2026-06-27 23:37Z

Gas Station, Bahia, Brazil, 1976 - Jay Maisel

illiquid @lefttailguy 2026-06-27 20:00Z

Dawg, you clearly have never been to Silicon Valley. 99% of people there (who are ultra wealthy by any reasonable standard, and certainly more wealthy than finance bros in Connecticut) have no public persona whatsoever. You do not know the names of most people who live in

Architectural Art (A-A) @4AAAAart 2026-06-27 07:40Z

Haara Metsa chandelier by Cameron Design House

Brandur @brandur 2026-06-27 22:06Z

Bought a new TV. I think this might be my favorite change since my last gen — a solar panel. You press a button on one of these things twice a day, and somehow that necessitated chasing after AAA batteries once a year, often involving a dramatic, wasteful overorder on Amazon.

Lazarz @Laz4rz 2026-06-27 07:26Z

Nvm, did a whoopsie and was sampling at 1/4 generation length, which is a good sign tbh cause it means you can generate 4 tokens at a time at some ~70% accuracy

Andrew Curran @AndrewCurran_ 2026-06-26 23:25Z

Semafor is reporting that The US government has lifted its block on Mythos 5 in a letter this afternoon from Commerce Secretary Howard Lutnick to Tom Brown. Fable is not included. This is US only, and covers 100 US institutions, including major companies and government agencies.

Jing Yu Koh @kohjingyu 2026-06-27 23:13Z

>Why has progress on computer use been so slow? I actually think progress has picked up a ton this year. The frontier models (anything past Opus 4.6 / GPT-5.4) are actually usable now. In 2025 none of the models were really reasonable beyond very simple tasks. >Computer use is

Susan Zhang @suchenzang 2026-06-27 01:47Z

this is how you grow to resent all the lowly irresponsible clueless plebs outside when you generalize this dynamic to "everyone out there wants to exploit us" inbreeding to protect themselves within their own bubble afterwards then becomes the only tolerable solution otoh

akira @realmcore_ 2026-06-27 20:53Z

I am currently doing what would've taken a week in like a day probably Essentially, architectural decision making, codebase crafting, end review/signoff etc. should be your primary tasks now Eval design is probably one of the most valuable skills in the world right now

GEOFF @geoffreywoo 2026-06-27 19:01Z

the next giant tech company may look like a spreadsheet crime scene for 18 months. not because the founder lacks taste. because the buyer pain lives in cells G through M, a pdf export, and a coordinator named lisa who knows which exceptions ruin friday. venture keeps asking

Charlie O'Neill @oneill_c 2026-06-27 23:24Z

An interesting question to think about is whether the labs will continual learn one big LLM or have personalised ones. If you did try and update fable with the experience of all users, it’s obvious you’d have to compress and throw away a lot of information. But how much? How much

Peyman Milanfar @docmilanfar 2026-06-27 23:53Z

Internal bodily rhythms (heart rate, breath) are not background noise for the brain The visual cortex seems to respond less strongly to visual stimuli that occur exactly when the heart contracts. The heartbeat competes with incoming visual information for the brain's attention

Max Hodak @maxhodak_ 2026-06-27 16:47Z

fun fact, political spam emails are perfectly legal in the US, and there doesn't even need to be an unsubscribe link. in fact, many systems treat "unsubscribe" link clicks as engagement signals!

“paula” @paularambles 2026-06-27 13:51Z

no, some of us grew up in cults and saw their childhood friends commit suicide after leaving

Pierfrancesco Beneventano @PierBeneventano 2026-06-27 20:27Z

High quality idea to high quality paper takes 10^3 prompts as of now. With PoggioAI/MSc we made it within 10. We do not need “full autonomous science” per se: We can hold the quality the highest and automate step-by-step.

Andrew Curran @AndrewCurran_ 2026-06-27 17:49Z

There has been a great deal of speculation about why Anthropic is keeping Mythos in restricted release. One of the least-discussed reasons is cost. Not the cost to Anthropic of serving the model, but the downstream effects that cost will have on the industry, and on the world.

zan @xenoaesthetics 2026-06-27 23:14Z

perhaps not only can one can engage in sequential batch processing of aesthetically opposed tasks, it may be a requirement, in perpetuity, for making progress apparently Caesar wrote philosophy of linguistics while spears flew on the battlefield, there is a deep unity beyond the

Arpit Gupta @arpitrage 2026-06-27 19:46Z

This really shows the role of subway lines in gentrification. Middle Village remains just outside the gentrification belt, but gentrifiers are moving through the L;J/Z;A/C lines so the frontier is now Glendale, Liberty Park, Cypress Hill etc.

Bloomberg @business 2026-06-26 22:38Z

SpaceX and Charter have held executive-level talks about partnering on a consumer mobile phone offering, according to sources

Aseel Swaid @aseelswaid9 2026-06-27 20:10Z

Lalibela, Ethiopia. Boaz Rottem, 2006.

fabiano.sol @FabianoSolana 2026-06-27 09:40Z

Arthur Hayes ( @CryptoHayes ) strikes again: 1. Teases to shill a "shitcoin" 2. Hayes' family office fund predicts $CARDS to $4 3. Sends $2M in $CARDS to a market maker 4.

47fucb4r8curb4fc8f8r4bfic8r @47fucb4r8c69323 2026-06-26 02:35Z

If we look at the percent of unhoused homeless as a percent of the US total population, we see a shocking stability across time and our current homeless population is returning to the pre-Millennial dip of the 2010s. That makes me wonder if there is a kind of min-max

Teddy @WarnerTeddy 2026-06-27 21:38Z

i dialed in reconstruction of the position of our balloon during it's latest ping even further using the dawn line at the time of our ping, the region covered by the ground station network we pinged traced via its packet broker forwarder hint and the payload's spreading factor

Tenobrus @tenobrus 2026-06-26 18:22Z

btw if u thought ur custom fanned-out data processing multi-agent orchestration harness was a particularly relevant long term moat, sorry, openai just added it as an api parameter

Michael @michael_chomsky 2026-06-27 23:43Z

I shipped an agent for a product with hundreds of thousands of MAUs last week. It had < 20 messages sent a day. Yeah. Nobody wants to use your agent.

Dylan Patel @dylan522p 2026-06-27 06:17Z

It literally doesn't matter? CXMT cannot satisfy China demand this decade. So there's 0nreason besides the small price arbitrage. It's like bannin Iranian oil. India and China buy it anyways, but price diff is less

will brown @willccbb 2026-06-27 09:13Z

pd disagg + wide-ep + kv offloading + cache-aware routing goes a long way

Nathan Lambert @natolambert 2026-06-27 15:28Z

Anthropic's political pressure on distillation is regulatory capture and most of the employees are blind to it under their veil of safety.

Basil Halperin @BasilHalperin 2026-06-27 16:33Z

Relatedly -- my favorite anti-monopsony bait: If sticky wages are via monopsony (rather than monopolistic unions as usual), then, *tightening monetary policy **raises** employment & output*... (for the same reason that a higher minimum wage boosts employment under monopsony)

arian ghashghai @arian_ghashghai 2026-06-27 23:07Z

imo as a rule of thumb when you see patterned disingenuous (or embellished) founder behaviors in a fundraising process it’s usually widespread VC behaviors that encourage it wrt to title inflation, VCs (wrongly) believe they can deterministically find early-stage investments

varun @varunneal 2026-06-27 23:12Z

Nesterov form seems a bit broken (m_t not defined) + you may want to rearrange to match the Lion-K/Muon style of nesterov thats nice with autodiff

Rishabh Agarwal @agarwl_ 2026-06-27 23:05Z

Problems we care about are often very slow verification loops (e.g, automating pretraining, making a new material)-- either you find a good enough proxy (e.g, simulator, grindable env) or unlock how to deal with this slowness (e.g, very sample efficient RL), which would be a step

Dwarkesh Patel @dwarkesh_sp 2026-06-25 23:57Z

Conquest's Third Law says that you should model any bureaucratic organisation as if it was controlled by a cabal of its enemies. It's a fun exaggeration, but it gets at an important phenomenon. Organisations have continuous generational turnover. The founding members eventually

elie @eliebakouch 2026-06-27 06:52Z

new inference optimization method by @deepseek_ai with an extremely detailed paper, draft model and framework to train them. results in production for dsv4 lead to +50% for throughput and latency (can go to ~80% for latency, crazy). full explanation of DSpark: it's about

xjdr @_xjdr 2026-06-27 19:10Z

we've been running GLM 5.2 in bf16 and in fp8 (experts and kvcache only, attention is always bf16) and have recorded virtually 0 measurable quality difference in our A/B tests and audits (very surprisingly) NVFP4 has shown a slight performance regression but could probably be

cinemoments @TheCineMoments 2026-06-27 13:07Z

9 shots tracking Mary's childhood home as the Backrooms keeps re-copying it... until all that's left is a blank yellow room. Memory, degraded one floor at a time.

altra @catboosted 2026-06-26 18:26Z

I’m consistently surprised how effectively domain expertise is able to be compressed into rubrics

Matt Slotnick @matt_slotnick 2026-06-27 19:09Z

how big does Anthropic need to become for Salesforce’s stake to represent 51% of its total enterprise value?

Kaito @kaycidot 2026-06-27 21:00Z

This also goes for all the new consumer fintechs using stables. What I'm seeing in successful stablecoin backed fintechs that actually get adoption: 1. Focus on a specific corridor - nailing localized onramps/offramps. ARQ, Belo, LemonCash, etc. 2. Focus on a specific target

Sebastian Herrera @SebasAHerrera 2026-06-27 14:08Z

This story is packed with exclusive details, such as some Microsoft staff being told in a March meeting that the future of software development means fewer people working harder by leveraging AI agents.

The Kobeissi Letter @KobeissiLetter 2026-06-27 17:04Z

US tech fund flows are seeing extreme volatility: US technology sector funds posted -$15 billion in outflows this week, the largest weekly withdrawal in at least 2.5 years. This exceeds the previous largest weekly outflow over this period by +100%. This also comes after +$19

Jay @jayair 2026-06-27 23:37Z

Some thoughts on DeepSeek Unlike everybody else, they are focused on reducing cost vs frontier intelligence; they are doing this by designing an architecture to run their models way more efficiently, and on Huawei's chips Their models are open source but only they seem to be

echen @echen 2026-06-25 17:51Z

we built HANDBOOK.md: a new benchmark for long-context agentic instruction following. we drop an agent into a live company environment with files (PDFs, Excel, Word Docs, ...), tools (email, Slack, Jira, calendar, ...), and a dense corporate handbook (up to 124 pages), across 5

David Klindt @klindt_david 2026-06-27 20:52Z

wait Matt Botvinick is also at Anthropic? Is anyone left at GDM?!?