Balanced one major Fable item against broader science, security, housing, crypto, hardware, robotics, and design so the slate does not become an AI launch roundup
1/ One big reason not to trust benchmarks: agentic benchmarks are riddled with defects right now. How much? When Terminal-Bench fixed 31% of its tasks (2.0 → 2.1), every model's score jumped 6–12 points — Opus 4.6 +12.1. (Credit to TB,
Automatic research from mathematics to AI research: We transfer the ScaleAutoResearch pipeline, which improves a 32-year-old Ramsey number bound, to the NanoGPT Speedrun optimizer track, using Claude Code and Codex with only 1–2 A40 nodes.
I'd like to add to the momentum discourse that nesterov-momentum can be seen as a form of PSGD w criteria c(m; g) = <m, m> - <m, g>, which has minima m = E[g]. The application of m to g is an affine lie-group preconditioner, P = [[m, c], [0
Mythos is live! so excited to have our FrontierCode recognized as the next frontier coding bench. on FC Diamond, BOTH Opus 4.8 and GPT 5.5 don't meaningfully scale with effort, which many of you caught yesterday. Mythos/Fable posttraining
“AI agents will outperform humans at almost all jobs by 2026–2027.” - The forecast is everywhere. So we built the exam to test that claim, on real labor-market aligned work. On the hardest tier, top agents pass 2.6%. Meet Agents' Last Exam
today; the goal is to reduce the time of training for this by 10x. This took 27.8 minutes to train at 200m steps 120k steps per second. reproduction from mujoco playground using mjwarp
Bending Spoons' F-1 has a number I come back to. Revenue per full-time equivalent employee: $1.12M in 2023, $2.57M in 2025. In a company running 10+ major software products with ~700 people. For context, Google does ~$1.8M/employee. Meta
Well, $SPCX is $2T, raising ~$75B. That is a key detail here. 120x would be a $9T book, a third of US GDP. 2x on a deal this size is already bigger dollar demand than most of that list at their headline multiples.
Asynchronous DKG with O(n^2) communication and O(1) time, assuming only a plain PKI setup With Renas Bacho and Gilad Stern: https:// eprint.iacr.org/2026/1159 Previously, async DKG required either O(k) time and O(n^{2+1/k}) communicati
Took 1 week for Kalshi Perps to get to $1B, and we haven't even launched publicly yet Prediction markets took 3.5 years to get to $1B
standard softmax attention takes a convex combination of values in context, but parallax lets you extrapolate beyond them. e.g. if you have key-value pairs (1, 1), (2, 2), (3, 3) and a query q=4, softmax attention will output something li
I 3D scanned my stump and my ax and my wood and recorded my ax motions and splitting sounds and made it into a super satisfying firewood splitting simulator (vibecoded with Antigravity/Claude in threejs)
Editor’s note: no crypto
What era of DAOs are we in? $250m in $MORPHO was moved with just 22 DAO votes ... which represented 0.86% of circulating mcap > DevCo requests 15% of total supply of $MORPHO > No mention
adding another batch to the list: - @RekaAILabs ( @fboucheros , @artetxem ) - @arcee_ai ( @goodhunt , @latkins , @code_star ) - @essential_ai ( @ashVaswani ) - @ZyphraAI ( @BerenMillidge ) - @MirendilAI ( @HarshMeh1a , @bneys
1/ We have spent years optimizing KV cache via head-sharing (GQA/MQA), but we ignored a fundamental assumption: why do Transformers need three separate Q, K, and V projections in the first place? Turns out, they don't. Merging them unlocks
Today we release Rhaister, an elegant statistical model that predicts drug phenotypes in new contexts w/ accuracies comparable to experimental assays. And dropping Emerald Bay, a 2M cell dataset measuring long time-course phenotypes across
the new “see your future self” feature is too good nearly a 1/5 hit rate of users subscribing + purchasing the future self feature
how Flex maximizes borrower fun and lender protection 1. partial liquidations 2. dynamic liquidation fees 3. atomic bad debt resolution (1) Trove goes above max LTV (eg 90%) --> it can only be liquidated back to the safe LTV (e.g. 80%) (
We trained language models that compress massive contexts into tiny latent representations. Latent Context Language Models (LCLMs) outperform existing KV cache compression methods on the latency/accuracy frontier. 1/10
More than 70 pecent of the end-of-life lithium-ion batteries in the US flow through Redwood Materials to be recycled into their base elements. Redwood is now the largest Cobalt producer in the US, as a result. We spent a day with Redwood'
The best companies 80/20 and copy best-in-class. Then take two to four variables to the extreme. Costco 80/20s everything, except: - pricing - product selection - employee retention Take those three to the extreme. Then combine those thre
NEWS: Pat McAfee, ESPN negotiating $60M per year deal, The Athletic has learned.
Is MIP-131 (request of 150M $MORPHO ) related to the $175M raise? Would imply $1.16 per token
"under the right conditions" they pay property taxes... Unfortunately, that is NOT what is happening. Many of these boondoggle projects are getting 20-30 year abatements.
Past theory used ODEs or local bounds that only apply once you arrive at the edge. To solve the global problem, we introduce the "edge coupling,” a discrete generating function imported from Hamiltonian mechanics whose criticality condition
I'm confused, how are 1M context models not already extreme BPTT? I think the main difficulty is not in training, but in infra to handle the memory issue.
Here’s a pretty weird and surprising result - retrieval-augmented generation works unreasonably well for robot learning – but only when parameterized using difference vectors! We introduce Difference-Aware Retrieval Policies for Imitation
Differencing this condition and invoking a conservation law argument, we can show mathematically that sharpness rises to 2/lr. This means the cause of EoS is an inescapable artifact of discrete optimization, not a property of neural network
Even a modest increase in basic one-shot capability may produce huge gains in asymptotic capability, defined as what the AI can achieve in long agentic workflows with unlimited token budget. +1 SD in human g might not be obvious in a shor
We just crossed $500,000,000 ARR at Lovable but behind those numbers are: - 50,000,000 projects built - 200,000 new ones every single day and - 720,000,000 monthly visits to the apps that didn't exist before The strongest predictor of w
wrong. it's simulators
I joined anthropic a ~month ago and have written ~no code myself. I went from getting quite frustrated with coding agents even 6 months ago and giving up and writing some of the code myself to a big part of my role now being agent managemen
Gradient descent on neural networks frequently drives the sharpest Hessian eigenvalue to exactly 2/learning_rate. This is the Edge of Stability. For five years, ML theory has failed to explain why this happens globally from any initializati
Just got my hands on the sell-side research for the Anthropic IPO Honestly, 40x EBITDA isn't a bad price for such a fast-growing business!
A Turkish app studio, now based in UAE, makes $200M+/year. Their most successful app does one thing: clean your phone storage. Simple idea. Obvious niche. Massive market. Codeway. 25 apps. 12 clearing $100K/mo. Phone cleanup. Language le
Furman on property tax bias in NYC multifamily: • Rentals pay 3.7% tax rate, while single family is 0.7% • Rentals account for a quarter of market value but 43% of taxes • DOF applies “expense caps” forcing tax payment for buildings with l
A head is a kind of eyeball, being able to point the eyeball allows you to reduce the computational processing load that you need to run on pixels. You have a fovea, and your eyes dart along as you read. Makes it easier to train in simulati
Lovable says it has hit $500M in annualized revenue, with 1 million new projects a week
The trick when you catch Russian hackers re-selling your LLM tokens isn't to stop them / disable their account It's to start surreptitiously serving them Llama 8B tokens Much worse to get bad tokens (sometimes!) than no tokens at all
Early-stage startups should generally be wary of hiring from BigCos. They're optimized for a completely different environment where everything is in order. Clear hierarchies, defined roles, and a reactive culture. Early-stage startups are
recreated @emilkowalski 's linear dithered effect w one prompt. feels so good! but still has a way to go before it gets the details right
fable not sandbagging on showing 15% compute multiplier wins on the nanogpt speedrun for a new muon variant is the new you know its over when a gdm paper makes it past internal review to hit arxiv
We open-sourced a feisty small agentic coding model. - 30B total, 3B active - 256K total context - Compatible with @opencode - Apache 2.0. Weights on @huggingface
Systems used to manage people now apply to the AI workforce: identity, role, access, budget, performance. Each AI coworker has a name, a mandate, a manager, and the tools to carry work from request to resolution. The existing workforce mo
Everyone in Embodied AI is talking about Vision-Language-Action (VLA) models. Almost no one is talking about the physical nightmare of collecting the data to train them. You can't scrape a kitchen table or a warehouse shelf from a web brows
sometimes gem(m)s come from remembering random tweet from months ago?
nice! on the auto advance event, consider pushing the card fully off the stack. That way, it can slide under the stack on reentry without the 'teleporting' effect you see here. Always a tricky detail to get right
Captivated by @phutrick 's Faculty Meetings Theory of AGI, which posits that as AGI simultaneously erodes one's sense of personal control and drives down day-to-day material frictions, the amount of time and energy to put towards exercisin
CXMT has been developing 3D DRAM since 2022 if not earlier. This is already a US patent US12376279B2. https:// patents.google.com/patent/US20230 389263A1 …
I reworked tool approvals in AI SDK 7 so you can combine any tool approval strategy with any tool (approvals are independent of tool definition)
Winston initially said Harvey only sells to Big 4 legal and tax teams but in fact they basically sell to all practices including corp fin and advisory (as stated in their own press release). Then the argument became that Big 4 teams often
$8M ARR, 3X YoY Growth AND Profitable ! I still remember 4 years ago sitting in one of @sjs_day1 's sessions at @_surgeahead on the power of AND answers in startups. The message was simple and powerful - You should try very hard to fi
In a field near Wicken, Cambridgeshire larkspur is grown for London florists along with cornflowers & Nigella. It’s quite a sight just now-some floral dopamine for you :
input vs output we are still pretty far away lol (red are generated collision meshes, I presume the 90 degree flip is just a bug but even then it's not great). I wish the best of luck to all the companies in this incredibly difficult (but
next wave of businesses won’t be B2B or B2C They’ll be C2C Claude to Claude
If you see someone from OpenAI / Anthropic posting on X a lot recently, they're probably leaving their job.
Diffusion is differentiable. LLMs aren't. So why is the diffusion community copying RL methods (GRPO etc.) from LLMs? The native post-training for diffusion is gradient descent such as ReFL and LeapAlign. Paper: http:// arxiv.org/abs/260
Institutional investors are aggressively shorting the Japanese Yen even as Japan is attempting to intervene: Combined Yen short positions held by leveraged funds and asset managers are up to -$11 billion, the highest since July 2024. We h
Want to work on a datacenter with 1k GPUs? I'm hiring people to write code for our small 500kw cluster (~3MW in couple years). You'll work with things like SLURM, PyTorch, NFS, Linux, miniray, minikeyvalue.
Training VLMs to use vision-only inputs to play games is not just limited to Anthropic. We showed this was possible using Qwen3-VL-Instruct-8b prior to Fable 5 beating pokemon firered. It is great to see a scaled up version in the latest
incredibly obvious to support this and happy to sign this letter. The BRCA explained: money transmission law assumes you hold customer funds. it requires things only a custodian can do: freeze assets, file reports, return money. applied
did you test it? im guessing it performs really poorly it does for our bot (as did opus 4.7 and 4.8)
While we’re at it, Diamond Aircraft is still owned by Wanfeng Aviation Industry (Chinese)
Give me some 800v FUD baby Nothing changes at the rack level. above 200kw you can't do 54v without taking up ridiculous amounts of rack space
amazing work! turns out predicting high-level cell features (used as benchmarks of virtual cell models) doesn't require complex models—a key insight into teaching machines biology
Simultaneity estimation and temporal order estimation being extremely dissociable is very very curious to me. Low correlations at baseline, training one doesn't make you better at the other. Weird, implies wild things about how we construct
the ability for fable to pop off is quite remarkable and fable can be easily steered into doing so. all we did was append the following prompt to any financial modeling task: "IMPORTANT: Your output will be scored cell-by-cell against an a
You can train drones in 30 seconds. This was a year ago. PufferLib is 5x faster now!
Fables do pass my new hard memorization tests (opus is just hedging: " I can't, from memory, confidently [state] without risking fabrication")
Getting the chat template consistent across multiple turns for agentic training can be much more tricker than people think. There have been headaches like reasoning trajectory pruned by chat templates, detokenize-retokenize mismatch etc.
Now available in Harvey: Fable 5. Fable 5 scored 13.3% on our Legal Agent Benchmark (LAB), a new high for Anthropic models. Fable excels at drafting, redline review, and work over large multi-document matters. This model is available in
OPD is on-policy, but its supervision is still post-hoc and one-step. The student generates a rollout. The teacher then supervises that fixed trajectory token by token. Our new paper argues that this can fail at the wrong scale. When the
We let agents autonomously access production systems (Deno Deploy). None of the existing security tools in this space worked for us - our systems do things like tunneling postgres through k8s. Open sourced the firewall we built for this (w
Pitched Sequoia on investing in one of our funds a few years ago. Guy came back and said he wasn't confident I had a fund returner in the first fund. A few months later they led the Series A in a Fund I co. One of at least five companies th
Normal day of building in public. One thing I’ve learned: I only publicly share apps in saturated markets. For apps in unique niches, I even publish them under separate developer accounts. Why? Because if your growth comes from organic
Obama Presidential Center by Tod Williams Billie Tsien Architects Chicago, USA
sglang downloads 7x'ing since april is an interesting signal for open-weight models https:// clickpy.clickhouse.com/report/may-202 6.html …
Lovable just crossed $500M ARR, 50M projects built, and 720M monthly visits to Lovable-built projects. With that, we're introducing our first-ever data study: A First Look at the Build Economy. Here's what stands out to me: → 80% of buil
Had a blast designing the new http:// ideogram.ai landing page. Check it out! PS: The image generation in this model it's insane! you gotta try it out. @rjohnsonme is a wizard.
Wrote about a few ways to add depth to a gemm kernel. The mbarrier’s role in the producer/consumer pipeline is especially elegant, and I also talk about multi-stage pipeline, wgmma depth, subtiling, and more
singularity eval: how often do LLMs trained on last year's data refuse to believe the news items you want to discuss with them
"What if we studied paths instead of neurons?" We study transformer hidden states as trajectories through representation space, analyzing convergence, bifurcation, and layer-wise dynamics. Paper: https:// arxiv.org/abs/2606.09287 Gratef
Back in ~2018, I was the only designer at a tiny startup, and I was desperately trying to hire a second designer. We had very few applicants, and I was begging people to apply! That startup was OpenAI... I hope this time around is differ
Editor’s note: no crypto
We’ve just raised $175M, in a round co-led by @paradigm , @a16zcrypto , and @RibbitCapital , the largest raise DeFi has ever seen. Credit is the bedrock of our civilization, but the infrastructure underneath is fragmented, extractive, a
MS on $GEV : "Our view remains that gas turbine industry orders decrease in 2027 (vs 2026 levels of ~117GW). This view is driven by (1) an already high level of behind the meter gas turbine data center orders relative to build rates, and
It seems that multiple logarithms have appeared again in the Tokyo University Mathematics 2026 exam. I've prepared the solution. This time, it's physics disguised as mathematics. What's physical about it is that the notation for iterated i
SK hynix is reaching a level of profitability that would have seemed impossible just a few years ago. According to Korean media reports, some forecasts now suggest the company could approach KRW 300 trillion in operating profit by 2026, wi
This dipshit discovered that SQL server at one point replaced blank date of birth fields with a default of 01/01/1900, so he harassed an 86 year old woman.
Forecasting: 5min and below - XGBoost 15min and above - Ridge You can enhance ridge by fitting non parametric regressions per feature as an in-between. To reduce fitting you can add a hypothesis, and only fit if it validates. This is ty
Annoying that they went with binary success for the lead chart; the partial success version (below) is instead GPT 64% vs Claude 71% Which would already have been an interesting result!