Backscroll ranked feed — 2026-06-27 23:10Z

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

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.

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.

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

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

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

Haara Metsa chandelier by Cameron Design House

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

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

Aseel Swaid @aseelswaid9 2026-06-27 21:38Z

Incredible photo. A woman sits in the ruins of her home in Tehran. Photographer Maryam Saeedpoor.

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

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

Lalibela, Ethiopia. Boaz Rottem, 2006.

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

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

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

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

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

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

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

dax @thdxr 2026-06-27 17:38Z

the models are very capable. i'm not saying they're very smart. but the bottleneck with them is likely your own imagination proof is so many companies demo their ai product with "it can schedule stuff for you"

Arpit Gupta @arpitrage 2026-06-27 16:37Z

The landlord reps sometimes claim that low rent on stabilized buildings causes higher rent in market rate, but this is the more economically sound mechanism. Stabilization lengthens tenure which lowers vacancy, tightening the market rate stock

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!

Just Another Pod Guy @TMTLongShort 2026-06-26 14:26Z

If you understand corporate America it is fairly intuitive that for the longest time the shortest path to managerial progression was increasing headcount in your department. Every cycle you made the case for why you needed more bodies as you executed against some sort of

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

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)

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

Alexander Doria @Dorialexander 2026-06-27 09:08Z

Yeah vllm/sglang do work quite well know, batching ensures a good theoretical throughout but then everyone want to Claude Code and you have to manage X parallel sessions, each with varying latency.

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

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

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

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

Aseel Swaid @aseelswaid9 2026-06-27 20:58Z

San Francisco shoreline 1957

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

Rahul Mathur @Rahul_J_Mathur 2026-06-27 04:43Z

Alienkind just raised a ₹30 crore round - this company was widely dismissed by observers as “Indiranagar PMF” since its 1st store went viral on Instagram in March 2025. This round was was at a ₹250 crore post money valuation (almost x3 the first round which was at ₹80 crore)

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

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

Alex Kesin @alexkesin 2026-06-27 19:48Z

This tracks -- something I've noticed though (at least on the VC-backed side) is that these ex China NewCos are heavily concentrated around Series A/launch rounds. China-origin has been ~1/3 of Series A capital in 2026 vs ~0% at Series C+ YTD

illiquid @lefttailguy 2026-06-26 16:03Z

AI slop as straussian infrastructure

Tenobrus @tenobrus 2026-06-25 22:45Z

it's not necessarily a victory for safety in the more absolutist unaligned ASI risk sense. but i do think these regulatory actions are probably safety victories wrt bio and cyber and general short-term mere megadeaths and societal disruption risk. if the USG has woken up enough

Soren Larson @hypersoren 2026-06-27 00:19Z

labs have to raise capital with incomplete information on attendant task complexity and hope they’re able to capture surplus sufficient to keep the ball rolling the application layer directly observes task complexity and matches the dumbest model sufficient for the application https:// x.com/mtslive/status /mtslive/status/2070652308925595920 …

Zafir Stojanovski @zafstojano 2026-06-27 16:19Z

Added this neat little figure to @natolambert 's RLHF Book showing the conditions under which we backprop through the PPO-style loss: 1. [Green region] Advantage is positive and π_new(a) < (1+ε) * π_old(a) -- the action we performed was good, so boost the probability of the

Bluechip @bluechip_org 2026-06-27 08:26Z

Binance is suspending EU services after failing to get one MiCA license. Bitpanda has three. That's one way of creating a moat. Exchanges that over-invested in compliance now have a market that rivals legally can't enter. Thomas Pecha from @Bitpanda explains at Bluechip25.

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.

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

em m0shouris @emm0sh 2026-06-27 16:05Z

i propose a new term: hard engineering, akin to the hard sciences it means engineering in the disciplines of mechanical, civil, electrical — anything touching atoms “i’m hiring hard engineers” is much easier than “i’m hiring engineers (no software please)”

Owen West @OwenWest91 2026-06-27 15:17Z

Worldwide armies are in various stages of drone doctrine. Prussians pioneered the bolt action rifle in 1871 but it took 25 years for the U.S. to adopt, thanks largely to retired veterans. Today our defense tech entrepreneurs often team with vets.

Dillon Mulroy @dillon_mulroy 2026-06-26 14:09Z

i suspect if the cost of SOTA models don’t come down and open weight models continue to progress, we’ll see the use of SOTA models reserved to efforts debated and decided via committee or RFC processes within enterprises and engineering orgs

Peter Kazanjy @Kazanjy 2026-06-27 20:59Z

Founders: Map every discovery question to product capabilities: Pain Discovery: 'What's biggest time suck?' → 'Deprovisioning takes 5-10 min each' Pain Sizing: 'How often?' → '10x daily = 2 hours/day lost' Demo Anchor: Show automation feature that saves exactly those 2

Architectural Art (A-A) @4AAAAart 2026-06-27 07:08Z

"Pixel Theory" modular furniture collection by Ceren Arslan, founder of the design studio EXIT

Matthew Cappucci @MatthewCappucci 2026-06-27 15:49Z

Have weather forecasts seemed less accurate lately? There's a major contributing factor: nearly half the morning weather balloons in the Lower 48 are "missing." This is an ongoing crisis that is degrading critical severe weather forecasts that we all rely on. It's having real,

Amir Efrati @amir 2026-06-27 14:58Z

? Anthropic’s big Slack bot launch, Claude Tag, sparked confusion / concern among employees of Salesforce, which owns Slack and promoted the Anthropic launch. Fox in the hen house?

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

waifupika @waifupika 2026-06-27 21:39Z

labs need to make auto router via api. all these tokens efficiency / harness default posts are unnecessary and irrelevant vs the core fix

Aurel Arnold @aurel_arnold 2026-06-27 13:41Z

Does collecting human intervention data (HG-DAgger) actually beat just collecting more demonstrations? I tested this myself: - interventions reached 72% success - more demos plateaued at 38% success I wrote up the experiments:

Nando de Freitas @NandoDF 2026-06-27 16:53Z

A model trained for next-token prediction is forced to build compressed representations of latent structure in text. Ilya Sutskever correctly refers to this phenomenon as understanding. Here, a model trained for next-step sensor prediction, with a robot that has proprioception

Charlie Bilello @charliebilello 2026-06-27 19:04Z

Bitcoin (-31%) and Gold (-6%) are the two worst performing major assets so far in 2026. This is something we haven't seen before in any calendar year.

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.

Andrew Curran @AndrewCurran_ 2026-06-25 20:34Z

The US Government has requested a slow staggered rollout of GPT-5.6, and OpenAI has agreed. During this phase the government will approve each user individually. This will probably be the norm for all frontier models from all labs from now on.

elvis @omarsar0 2026-06-27 18:49Z

If you use LLM-as-judge, this one is worth reading. (bookmark it) It's actually one of the most effective ways to use LLM-as-a-Judge for evals. Holistic judge scores hide both their reasoning and their ceiling effects. BINEVAL decomposes each evaluation criterion into atomic

Florian Brand @xeophon 2026-06-27 05:30Z

building verifiers has shown time and time again how important ownership of your harness is good luck trying to patch closed source harnesses to make them more useful to train or eval

Ryan Brewer @ryanbrewer 2026-06-27 19:39Z

Pretty funny disparity from how I view computer use. A year ago it was completely useless, and now its gotten so good that I can actually offload a lot of meaningful work to it

junkbondinvestor @junkbondinvest 2026-06-26 20:44Z

Apollo, a large manager, put out a chart showing deep distress is concentrating in small private credit funds. Loans marked below 50: ~12.5% at small funds, ~8% at big ones

Dmytro Dzhulgakov @dzhulgakov 2026-06-27 17:31Z

DSpark from @deepseek_ai ingeniously integrates many speculative decoding ideas to achieve 1.5x to 5x higher throughput in a real production system Let's understand it with 10 ideas, starting from the very basics

Ethan Gao @ga0_lab 2026-06-27 15:00Z

Did you know you can create the Maxwell-Boltzmann distribution with just some balls and a motor? I built this simple device to illustrate it. This is how it works: A spinning agitator pumps energy into ~400 balls, so they bounce around and collide like gas particles in a