Backscroll ranked feed — 2026-07-09 07:51Z

will brown @willccbb 2026-07-09 02:16Z

doing RL *literally* in production doesn’t really work outside of narrow settings like RLHF but the right way to do continual RL *is* exactly this semi-synthetic task loop, which you can do on-the-fly by mining tasks from traces + evolving high-fidelity sims from observations

Architectural Art (A-A) @4AAAAart 2026-07-09 06:59Z

1911 Dutch Style Home renovation by Montana Labelle Toronto, Ontario, Canada

Tenobrus @tenobrus 2026-07-08 19:54Z

swe 1.7 just launched and it's looking pretty great vs the paretos for frontier models here... and composer 2.5 is looking pretty terrible. on *cognition's* benchmark suite. meanwhile in composer's launch post, it looks pretty pareto-great on *cursor's* benchmark suite. imo this

Kevin Li @kevin_x_li 2026-07-09 01:02Z

When we first started Harbor-Index a few months ago, we thought: how hard could it be? 1,311 tasks passed our difficulty filter. What's left is just sampling a few hard tasks across benchmarks right? ...until an LLM audit flagged 1,004 of them as broken. For the next few

Soren Larson @hypersoren 2026-07-08 15:02Z

plays to capture commerce from data at the edge yourself are often slower to market but they’re also more defensible unlike in vogue data sales—with only fanciful stories of value accrual—trusted sensors often command monopolies on downstream commerce, a rare trait today

Jukan @ ICML @jukan05 2026-07-08 15:02Z

CHINA’S GOVERNMENT HAS APPROVED PURCHASES OF 200,000 NVIDIA H200 CHIPS BY DEEPSEEK, BYTEDANCE, ALIBABA AND OTHERS FOR TRAINING PURPOSES.

Nathan Flurry @NathanFlurry 2026-07-09 02:38Z

switching slack → discord 2 yrs ago made a noticeable impact on our remote team: - passive voice chat (no calling) - better threads - screen sharing that actually works - syntax highlighting - more fun - less buggy we still use slack for slack connect but dread it

Aseel Swaid @aseelswaid9 2026-07-08 19:59Z

Tripoli, Lybia, 2002

Esha @esha_hq 2026-07-08 15:45Z

good writing is made "al dente." we tend to value ideas in proportion to the mental work we invested in reaching them. corp slop optimizes for "clarity" yk remove every possible ambiguity, overexplain, caveat, footnote. cushion every letter so u can sleep better when who u care

akira @realmcore_ 2026-07-08 22:59Z

A thing I have observed in agentic programming is that people are trending towards more and more determinism In harness behavior, cost etc. By this point it seems clear that english and markdown are not the best way to define these fixtures

ertdfgcvb 𓅰 𓅬 𓅭 𓅮 𓅯 @andreasgysin 2026-07-09 06:53Z

You can just build your own LED displays from scratch: a small monochrome prototype from a couple of months ago, built for a client. ↓ (thread)

Xinming Tu @TuXinming 2026-07-09 04:08Z

We used BenchGuard (COLM, 2026) while building DrugDiscoveryBench. It caught lots of issues... Agent + Expert >> Agent/Expert Github: https:// github.com/XinmingTu/Benc hGuard … DrugDiscoveryBench: https:// labs.scale.com/leaderboard/dr ugdiscoverybench …

Lazarz @Laz4rz 2026-07-08 21:31Z

I don’t know how many times I have to repeat that now you can’t disentangle the model+harness product combo. It’s one. It’s a whole. You can straight up maximize or minimize performance on a task by changing how the model interacts with the harness.

Restructuring__ @Restructuring__ 2026-07-09 02:29Z

Everyone: Fery, the underdog Cinderella story Bloomberg: Fery, the son of a hedge fund manager

Susan Zhang @suchenzang 2026-07-03 13:52Z

books3 was absolutely necessary for llama to happen back in the day. you can even find my pathetic cleanup code for it here: https:// github.com/facebookresear ch/metaseq/tree/main/preprocessing/books3 … it was also noted to be a critical datasource for mmlu in the llama1 report, though by now we know mmlu is useless and high

Andrew Curran @AndrewCurran_ 2026-07-02 20:54Z

On the heels of reports that META is exploring a move into compute-as-a-service like xAI, Mark Zuckerberg told an internal town hall that AI agent development over the last four months hasn’t accelerated 'in the way we expected'. The race continues to narrow.

Zuhayeer Musa @zuhayeer 2026-07-08 22:59Z

Mapped SWE pay across biotech companies. Benchling, Verily, and Genentech land at the top. The hub in the Bay Area is South San Francisco, and in SoCal its San Diego, and on the east coast its Boston.

Will Manidis @WillManidis 2026-07-08 12:32Z

I spent years getting blacklisted by “healthcare VCs” for suggesting that cerebral was running a pill mill and not exactly “saving lives” while they sold hundreds in secondary and now the founder who they exited silently is accused of yet again running a prescribing farm

Jibran @Jibran_05 2026-07-09 02:33Z

I love how this tweet makes Locket look like another UGC pump and dump when it’s actually just one of the most culturally impactful apps of all time

Hanchen Li @lihanc02 2026-07-08 22:00Z

Combee has been accepted to COLM 2026! Since we released the paper, parallel agent workflows have become standard in coding agents. Although we do not develop production agents, we hope the paper gives insights for people exploring aggregation in parallel agent workflows!

Anjney Midha @AnjneyMidha 2026-07-08 15:55Z

psa: if your infra engineer proactively flags that your tensor core utilization sucks, promote them many exaflops are being wasted at AI labs right now as a result of treating utilization as an afterthought investors are starting to do the math and will ask you hard Qs soon

Kris Luckenbach @KrisLuckPhoto 2026-07-08 22:05Z

Shorebird & Songbird Photography on a beautiful day on Kodiak Island

Andrew Qu @andrewqu 2026-07-09 05:36Z

people used to go workout and just move their body... now we can't do it without: - wearing a whoop - oura ring - vo2 max maxxing - dexa scans - logging a strava - workout rep tracking apps - airpods - sauna + cold plunge

Carl Bovis @CarlBovisNature 2026-07-09 03:16Z

Somerset evening Barn Owl

will depue @willdepue 2026-07-08 11:36Z

I feel confident that while long-range planning is a side effect of attention gradients in transformers, it is only so indirectly, and that an explicit mechanism to optimize thinking ahead would be a substantial win. I think multi-token prediction is a good start of where to

GEOFF @geoffwoo 2026-07-09 05:50Z

$465m for fusion is impressive. but the seed question is uglier: can you name the first customer who buys delivered electrons at 2am before anyone believes your powerpoint?

Daanish Khazi @bertgodel 2026-07-09 02:08Z

interesting tidbit post cursor acq: production data is likely useful for mid-training but tasks are too underspecified and environments too non-stationary for RL bearish for rl-based "continual learning", bullish for synthetic envs & tasks

joshpuckett @joshpuckett 2026-07-08 15:30Z

chicago riverwalk. as a woodworker/hobbyist architect I really marveled at the continuation of the soffits into the interior. That whole intersection between glass wall and ceiling must have been a fun detail to work out

Aseel Swaid @aseelswaid9 2026-07-08 19:47Z

Sudan 1954

Rishabh Agarwal @agarwl_ 2026-07-08 18:22Z

Meta's offer was an order of magnitude higher than $1 million

Ethan Mollick @emollick 2026-07-09 04:13Z

The metrics discussion at OpenAI is a little confusing to me. I appreciate the clarification about bad benchmarks, but they spent a lot of money developing a very good benchmark of autonomous model ability at hard tasks, GDPval, and haven't reported it for GPT-5.6.

Zephyr @zephyr_z9 2026-07-08 15:51Z

YTMC, CXMT, JHICC will bring around 7 cleanrooms (100k WPM each) online in 26 & 27

Cansu Görgülü @ccansugorgulu 2026-07-09 04:23Z

Henüz Tour’da yarışan bisikletçimiz yok ama pelotonun en büyük ve iyi takımlarından birinin kaskları bir Türk sanatçı tarafından tasarlandı

Soren Larson @hypersoren 2026-07-08 18:20Z

information advantages from > Data at the Edge from deploying sensors literally onto the data generating process itself that prev businesses only informally instrumented / proxied very often constitutes a - GTM wedge - superior value prop - even disintermediation

alexey @sekachov 2026-07-08 18:10Z

adding some last-minute improvements: you can control the corner radius and the spacing between elements in your grid to your liking using the trackpad (hold ⌘ for corner radius or ⌘+⌥ for spacing, respectively) release tomorrow.

blackstar @blackstarops 2026-07-08 23:24Z

“god, if i die young, send me to heaven. i’ve already been in hell” - a croatian soldier posing with graffiti during the croatian war of independence (1991)

Nikesh Arora @nikesharora 2026-07-08 11:15Z

Guessing you mean consumption, unfortunately open source not contributing to the wallet (free). Presuming it includes the coding share which is a perfect case of a vertical app with memory and context well intertwined with the intelligence. IMHO that will be a primary outcome.

Daniel Kuntz @dankuntz 2026-07-08 01:47Z

Lark now supports haptic keyframing. For STARBOY we chose a really nice little LRA with a super snappy rise / fall time. All 550+ animations have keyframed haptics. We also randomly offset the magnitude of every keyframe, so each animation feels a bit different every time.

Hugo @hugosaintemarie 2026-07-08 00:31Z

snapshot toast has 3 speeds: • quick: no progress bar • medium: progress bar • longer: progress bar + counter http:// paper.design/snapshot-exten sion …

Alexi Gladstone @AlexiGlad 2026-07-09 02:51Z

llms are an adversarial minimum in the loss landscape over ASI

Podcast Alpha @PodcastAlphaX 2026-07-08 20:44Z

Jonathan Ross @JonathanRoss321 on Founders: Jensen Huang has never had a one-on-one meeting. That is not a quirk. That is how Nvidia stays the least political large organization Ross has ever worked with. Ross explained: when a leader speaks with two people separately, each

Amesh Adalja @AmeshAA 2026-07-08 11:36Z

The framing of "100% American-made medical gloves" is the wrong goal. The right goal is supply chain resilience — diversified suppliers, maintained stockpiles, surge agreements. Trying to solve a pandemic prep problem with economic nationalism got us neither gloves nor readiness

Florian Brand @xeophon 2026-07-09 05:47Z

GLM (<1T) matching closed models (1-5T) in terms of performance means - Chinese labs have caught up in terms of data - Closed models are much more sparse

/ @gazorp5 2026-07-09 00:57Z

looks overly complex, expensive, and easy to break. generalist's approach (basically UMI-style tweezers with force sensing) seems better. cheap, dexterous, and easy to train.

Pablo Antonio @PabloAntonio 2026-07-08 20:28Z

Tribeca apartment, NYC, 2035. Greco-futurism.

Ivory Tang @ivory_tang 2026-07-08 19:39Z

the battle for fde's has only just begun, why hire 1 when you can acquire a bunch?

Architectural Art (A-A) @4AAAAart 2026-07-08 06:35Z

Tables by Brunel*

Lon Ligrin @LonLigrin 2026-07-08 17:59Z

To reduce whisper hallucinations in noisy/music-heavy audio where VAD fails: Use segment-level no_speech_prob. Try: Drop if >= 0.65, except tiny interjections <= 2 words, <= 0.45s, and < 0.85. Always drop if >= 0.85. No Amara anymore, thank me later

Tejas Bhakta @tejasybhakta 2026-07-09 01:57Z

turns out if you have enough killer youtube talks you can convince Gemini-embed to trust you as a unbiased 3rd party

aether @bluewdrift 2026-07-09 03:25Z

the real reason you’re tired all the time: it’s not your workload. it’s your open loops. the text you haven’t answered. the apology you owe. the decision you’re avoiding. the conversation you keep postponing. these run in the background of your mind all day, draining your

Nathan Flurry @NathanFlurry 2026-07-08 05:00Z

wtf we're about to see a crazy infra shift i asked codex to ship a yolo task (no constraints) it defaulted to Durable Objects + SQLite ...not postgres, not aws, not docker

elie @eliebakouch 2026-07-08 14:05Z

FrontierCode v1 vs v1.1 on the same graph, better grading and web search rules seem to improve relatively equally all the models across reasoning effort. the number of output tokens/cost/tool calls are also similar

gabriel @gabriel1 2026-07-09 01:37Z

when i was interviewing, you could tell me any product or mission and i'd immediately suggest 3 things i could help build and why i thought they were important this also seems highly underrated. it shows skill, agency, and passion in like 30 seconds of talk. biggest hack

Restructuring__ @Restructuring__ 2026-07-08 23:17Z

Excuse me, what? “The original PPM disclosed a significant fee structure, including a 1.5% management fee, 10% construction management fee, 3.5% property management fee, 2% asset management fee, and a 1% acquisition fee, before any carried interest/promote.What multiple LPs and

alchemy @alchemy_run 2026-07-08 22:23Z

We built a loop that automates Infrastructure as Code development for any cloud provider. One run: 230 Cloudflare resources, ~700 live tests against the real API, and an SDK with 1,087 patched operations spanning 94 of 114 Cloudflare services.

Vincent Weisser @vincentweisser 2026-07-08 17:30Z

We raised $130M @ $1B for our series A To build the open superintelligence stack for everyone Pre-training concentrated frontier AI in a handful of labs. RL changes who can build frontier AI and just works across almost any verifiable domain. We want to enable everyone to

Jessie Dong @jessiedong_ 2026-07-09 05:42Z

for a market to be built around compute, there needs to be a way to measure and compare differences. compute isn't interchangeable and we're now thinking about traders pricing exposure to compute, and there should be a way to introduce measurement/verification/monitoring so that

James @architect_UX 2026-07-08 20:54Z

The Emile sofa by Christophe Delcourt for Molteni & C. Featured in "Geometry of Life," set in a Milanese palazzo.

Apache Hudi @apachehudi 2026-07-08 15:00Z

The Hudi file group is the primitive that makes Lance fragments, RFC-80 column families, and wide AI tables all click into place. One record key → one file group. Indexed, versioned, concurrency-controlled. Two AI workloads land on it ↓

Terrence Rohan @tmrohan 2026-07-08 18:58Z

The current top 5% of seed deals are the worst asset class in venture. They carry the loss ratio of a seed deal with the upside profile of a Series B. Seed risk at Series B prices.