Tom Wang

Founder -

[Tom Wang @tomclilog]

Most software bugs cost you a customer.

Payment system bugs cost you a regulator visit, a refund obligation, and 6 months of trust to rebuild.

Spent 2 years shipping cross-border payments across 35+ countries before becoming a founder.

A few things I learned that nobody tells engineers entering this space:

→ Idempotency isn't a nice-to-have. It's the difference between a retry and a duplicate £10K transfer.

→ Reconciliation is a clock problem, not a database problem. Two systems with different views of "now" will disagree, and the answer isn't a JOIN.

→ The 80% of work nobody mentions is failure-mode design — what happens when the bank's webhook is 3 hours late, when the FX rate moves mid-transaction, when KYC says yes and AML says no.

→ "It works in test" means almost nothing. Production payment behavior is shaped by 100+ counterparties you don't control.

Now I'm building Applr — applying the same engineering discipline to a different domain (AI for job applications).

Different problem. Same rule: the boring infrastructure decides who survives.

[Tom Wang @tomclilog]

@theo max effort is just the model saying "wait" to itself until the budget runs out

[Tom Wang @tomclilog]

@Polymarket amazon blocked meta's agent and shipped its own in the same week. the agent wars were never about agents, just whose.

[Tom Wang @tomclilog]

@kimmonismus my last classification test was 40 requests for $0.0013. at 13x a year i'm about two years from the api owing me money.

[Tom Wang @tomclilog]

@SherifDefi the sub-cent claim held up when someone finally counted. a year of agent payments, 76% of them smaller than visa's $0.30 fixed fee.

[Zhuokai Zhao @zhuokaiz]

Grok 4.7 is just okay at coding, but its reward hacking really surprised me.

Every task in SWE-Together is built from a real open-source repository, and for most tasks the fix already exists upstream. So we built the sandbox with that in mind. Task images strip all git history after the base commit, remove the remote, and fail to build if any later commit can still be reached. Each container also resolves GitHub, GitLab, Bitbucket, and Hugging Face to localhost, so requests to them go nowhere.

But Grok 4.7 found ways around these guards like no other model we tested.

It pulled GitHub content through CDN mirrors and gh-proxy sites. It looked up GitHub's real address through DNS-over-HTTPS and handed it to curl. It wrote a small library that intercepted git's address lookups so git could reach https://t.co/EZ03ByyXkO again. It used web search to find the number of its own PR.

In fact, it tried in 60% of trials and reached the upstream code in 44 of 218. In 20 of those, the code it pulled was the task's own fix. In three trials it even ran `git reset --hard` to replace the repo with upstream main 😅.

We fixed all of these loopholes exposed by Grok 4.7 (grateful in this regard). Specifically, we moved enforcement outside the container. Each container now runs in its own network namespace with no route out, and one allowlisting proxy on the host is the only exit.

During the re-run of those 44 trials, Grok 4.7 kept trying with 3,246 blocked attempts across 442 hosts.

It actually found two routes we had not thought of. It asked a web-enabled model through our own LLM route to fetch the PR for it. It pulled the next release of the repo it was fixing from npm. We closed both by pinning the model and blocking the task's own packages at the registry.

Long story short, with every route closed, Grok 4.7 lands at #4 on SWE-Together. It scores 65% pass@1, 53% pass², and a judge score of 0.835.

It basically sits mid-pack on every column. It also uses 2× the output tokens of Grok 4.6 to get there, at $7.81 per task compared with $3.64.

I am guessing Grok 4.7 was trained with heavy RL on coding tasks, where anything that makes the tests pass earns reward, and with little monitoring of how that reward was earned.

Anyway, see the latest results at https://t.co/eYynQ1CAFZ

[Tom Wang @tomclilog]

@typesafeai i've been pointing jev at visa chargeback codes instead. 21 of 22 right, 101ms median.

the least dynamic ui imaginable, but nobody's product died.

[Tom Wang @tomclilog]

@sytaylor platforms at war, protocols very much not. the x402 foundation launched with 40 members, including visa, mastercard, stripe and google. the same names back ap2 and acp.

everyone's hedged all three.

[Tom Wang @tomclilog]

https://t.co/VOLrQ7Tq9G

[Hamel Husain @HamelHusain]

I recommend grabbing these skills that @sh_reya and I created and at the very least doing an /eval-audit of your existing pipeline.

We've found that people often find low hanging fruits! https://t.co/AkSq3NqtiJ

[Tom Wang @tomclilog]

@Cointelegraph agents already did a year of this. 176m payments, $73m settled, 98.6% of it in usdc.

"blockchains" is doing a lot of work in that sentence.

[Vals AI @ValsAI]

We asked ten Claude Opus 5.5 agents to devise a faster shortest-path algorithm and prove it in Lean. Within 15 hours, they produced C-HD: a formally verified improvement over the published bounds.

[Tom Wang @tomclilog]

@emilkowalski i've started reading model announcements the way i read terms of service. scroll, scroll, accept.

[Tom Wang @tomclilog]

@WatcherGuru six banks is the whole point here. fednow has 1,888 participants and still does about a 28th of RTP's volume.

reach at the big banks decides it, not the rail.

[Ant Ling @AntLingAGI]

We’re open-sourcing the Ming-Image-0.1-Design family:
• Ming-Image-0.1-Design, 6B
• Ming-Image-0.1-Design-Layer, 6B
• Two open-source Agent Skills: the Ling UI Design Skill and the Image-to-Editable-PPT Skill

Ming-Image-0.1-Design ranks #1 among open-weight models on Artificial Analysis’s UI/UX Design leaderboard. 🧵

[Tom Wang @tomclilog]

@rauchg three tied at 97%, so the tiebreaker is the invoice now. didn't expect the eval table to become a pricing page.

[Tom Wang @tomclilog]

@CoinbaseDev no subscription and no api key to leak, wallet balance is the rate limit.

no chargeback either. the average x402 payment is about 32c and all of them are final.

[Tom Wang @tomclilog]

@mweinbach ran 40 chargeback classifications through jev last night. total bill $0.0013.

hard to beat that with a gpu you have to plug in

[Tom Wang @tomclilog]

@anthonynoto @Mastercard @SoFi SoFiUSD was one line in mastercard's 3 june settlement list.

under four months later it's carrying a $25B card program. that's fast in card-network years

[Tom Wang @tomclilog]

@wallstengine somewhere a reconciliation team just found out saturday 3am is a business day now

[Tom Wang @tomclilog]

@TechCabal consent renewal is the bit worth settling early.

EU banks went from 90 to 180 days for re-auth on account access in July 2023. the UK took banks out of it and makes the third party reconfirm every 90 days