deepsy

5.2K posts

deepsy banner
deepsy

deepsy

@deepsydoin

God forbid white boy get a lil motion Building @alembiclabs

Katılım Mart 2022
399 Takip Edilen3.1K Takipçiler
deepsy
deepsy@deepsydoin·
woke up to find a huge red candle. 5+ wallets dumped overnight while i was asleep ~20 SOL out, MC fell to 14k. then morning saw the chart, assumed something was wrong, more selling. we're at 7k now. honest take: this doesn't change much for me. dev fees haven't been enough to buy real runway anyway, and the lab was already running on principle, not capital. what's still true: — the lab is still shipping folds (now on the post-audit fixes from yesterday) — the 3d interactive scene is still going out soon — grant applications are still in flight — the science doesn't care about the chart i'm not trying to spin this. PA dropped, that's real. what i can control is: keep building. keep shipping. let the work speak. if it ends up mattering, the people who matter will notice. if it doesn't, i'll know i did everything i could with what was in front of me. back to it. D65AmX9aCF3wY4F4iwcGAfMtTyabTiD3YDtaX4uLpump here ca if you want support $alembic and me
English
0
0
0
32
deepsy
deepsy@deepsydoin·
quick follow-up on yesterday's audit fixes. after deploying, post-fix folds came in with a too-high DISCARDED rate. turned out the new Structural prompt had an escape-hatch phrase ("known failure mode") that the LLM was over-applyin — discarding folds with strong metrics (pLDDT 0.78, ipTM 0.92) just because the modification used non-canonical chemistry. closed it with a 3-layer fix: — rewrote the verdict criteria with explicit "what is / is not a verdict signal" — unmodelled chemistry goes to caveats, never to verdict — added a deterministic metric floor in code. if pLDDT ≥ 0.70 and ipTM ≥ 0.40 with no downgrade signals, the LLM physically cannot discard. floor only lifts, never demotes. — retroactive reclassification of 8 folds that got over-discarded under the old prompt — all moved to PROMISING with a caveat note. running a few sanity-check cycles now. once those pass, the lab goes on full throughput for 24 hours. audit → fix → re-audit → fix again. that's what compounding looks like.
Alembic Labs@alembiclabs

audit on the audit. yesterday's verdict fixes were over-discarding folds with strong metrics. shipped today: — escape-hatch removed from Structural prompt — deterministic metric floor in code — 8 folds reclassified DISCARDED → PROMISING sanity cycles running. full throughput in a few hours.

English
1
0
2
69
deepsy
deepsy@deepsydoin·
today's plan for @alembiclabs : — submitting the next batch of funding applications — BIO Protocol launchpad, Molecule, VitaDAO. Vril already in. — continuing work on the 3d animated lab. agent animations going in today. — and a hard test — running the lab at full throughput for 24 hours straight. throttle off. see how it holds up at scale, what the output looks like, where the bottlenecks actually are. back to building.
deepsy@deepsydoin

Gm

English
5
3
7
511
deepsy
deepsy@deepsydoin·
Gm
QST
4
0
7
635
deepsy
deepsy@deepsydoin·
Gn peptiders
Indonesia
2
0
6
250
deepsy
deepsy@deepsydoin·
While the lab keeps shipping folds (now running on the new fixes from yesterday's audit),i'm finishing the 3d interactive lab. each of the 5 agents gets visualized as a working character — generating animations for researcher, literature, clinical, structural, communicator. you'll be able to walk into the scene and watch them work. functionally it changes nothing about the science. but no one in this space has built something like it, and i think it's going to be something. content engine, virality, immersive proof that "autonomous lab" isn't just a tagline. once shipped, plan is to livestream the scene on @pumpfun 24/7. hope I can ship it in next few days
English
3
1
19
1.3K
deepsy
deepsy@deepsydoin·
research log #1 — what 51 folds taught the lab. spent today auditing every fold the lab has produced since launch. 51 folds. 9 REFINED, 18 PROMISING, 24 DISCARDED. that ratio is roughly what i expected — discarded outnumbering refined is the system being honest about uncertainty, not a bug. but the audit surfaced six concrete patterns that need fixing. all six shipping today. — what the data showed PATTERN 1: short peptides break the tool.every Epitalon fold (5 of 5, AEDG = 4 residues) returned pLDDT < 0.35. four-residue tetrapeptides are below Boltz-2's resolution floor. the lab kept proposing Epitalon variants and burning compute on predictions it couldn't make. PATTERN 2: class B GPCRs + non-canonical residues = systematic failure.all 6 Sermorelin and Tesamorelin folds DISCARDED. AlphaFold-family models aren't trained on Aib, S5 staples, hexenoyl caps. @DelixLabs flagged this earlier — confirmed in our data. PATTERN 3: lipid and uncharacterized targets give zero binding signal. Selank × Tuftsin/NRP-1, SS-31 × Cardiolipin, DSIP × unknown — 6 folds with ipTM = 0.0. Boltz-2 doesn't model peptide-lipid interactions. shouldn't have been running these. PATTERN 4: Boltz-2 affinity prediction was sitting unused. binding_probability and binding_pic50 captured in 0 of 48 folds. the lab wasn't asking for them. PATTERN 5: UniProt resolution failing on 6 folds. Researcher → Clinical handoff dropped canonical target IDs — downstream ChEMBL queries fail, Clinical can't ground hypotheses in real bioactivity. PATTERN 6: same 14-section Communicator template for REFINED and DISCARDED. DISCARDED reports don't need mechanism deep-dives. they need tool-limit context. — what shipped all six fixes deployed today: 1. Boltz-2 affinity prediction wired into the structural agent with defensive fallback. discovered along the way that BioLM's affinity head currently targets protein-ligand pairs, not protein-protein. for peptide-receptor folds the binding columns stay NULL until we expand into ligand-target work. infrastructure ready when the modality fits. 2. target-resolution gate. lipid targets, putative receptors, missing UniProt IDs get refused before Boltz-2 runs. saves ~$1.50 per blocked fold. live-verified on folds #56 (SS-31 × P-glycoprotein) and #57 (DSIP × GABA-A) — both caught at the gate before structure prediction even fired. 3. tool-limit registry consulted by Researcher. peptides under 5aa, class B GPCRs with non-canonical residues, lipid/putative targets — all flagged. live-verified: fold #56 first proposed cardiolipin, gate fired, Researcher regenerated with a different target. 4. cross-fold memory penalty. peptides with 3+ consecutive DISCARDED and zero REFINED get added to an AVOID list — Researcher refuses to propose them. Epitalon, Sermorelin, Tesamorelin, FOXO4-DRI all auto-blocked from current cycles until tooling catches up. 5. Chai-1 agreement now feeds verdict logic. agreement < 0.40 = deterministic downgrade one tier (REFINED → PROMISING, PROMISING → DISCARDED). ensemble disagreement is signal. 6. branched Communicator templates. REFINED gets the full 14-section report. DISCARDED gets a 6-section template that's honest about tool limits — cuts ~25% of token cost on roughly half of all folds. fold #60 demonstrated it cleanly: TLDR opens with the discard reason, cross-references prior failures, identifies class B GPCRs as the limiting factor on its own. a few infrastructure patches landed alongside: — JSON parse retry on Communicator (the new templates are complex enough that occasional parse failures needed defensive handling) — database session recovery after Communicator failures (slug and on-chain steps were reading expired ORM attributes — fresh session + refetch fixed it) — idempotent DB migration for the new discard_reason column — takeaway real research compounds. the next 50 folds should be measurably better than the last 50 — fewer wasted cycles, sharper signal where confidence holds, less interpretive reach in the PROMISING bucket. a lab that doesn't audit itself is just a content generator. closing the loop between observation and improvement is what separates real research infrastructure from sophisticated text generation. audit data + dataset: alembic.bio/folds source: github.com/alembic-labs/a…
deepsy@deepsydoin

the $alembic lab has shipped 50+ folds. small dataset, but big enough to actually examine. going through all of them today. writing up patterns, failure modes, methodology gaps, surprises. shipping the architectural improvements that come out of it. a lab that doesn't audit itself is just a content generator. real research compounds. research log dropping later today.

English
1
1
9
1.5K
deepsy
deepsy@deepsydoin·
the $alembic lab has shipped 50+ folds. small dataset, but big enough to actually examine. going through all of them today. writing up patterns, failure modes, methodology gaps, surprises. shipping the architectural improvements that come out of it. a lab that doesn't audit itself is just a content generator. real research compounds. research log dropping later today.
English
1
0
7
898
deepsy
deepsy@deepsydoin·
👀👀👀
deepsy tweet media
QME
0
0
10
378
deepsy
deepsy@deepsydoin·
Warmlt thank everyone for support♥️ $alembic has a chance to become a real example — proof that a @Pumpfun token can grow into a legitimate company, attract serious investment, earn proper attention, and give back to the ecosystem that enabled it. if it works the way it's meant to, everyone wins. i'm grateful this community is giving me the chance to build something meaningful. i'll do everything i can with that opportunity, and i'll keep building in public the whole way through. back to cooking 🧑‍🍳
English
0
2
15
775
deepsy retweetledi
WhenBag
WhenBag@WhenBag·
D65AmX9aCF3wY4F4iwcGAfMtTyabTiD3YDtaX4uLpump 40 K
deepsy@deepsydoin

some thoughts on the DeSci ecosystem and where @alembiclabs fits in it. i've been thinking about this a lot since launch. the bio/acc and DeSci space is moving fast. more projects every week. real funding. real research. and a real temptation to frame everything as competition. i don't think it is. and i want to say that publicly. — the space is too big to compete in the underexplored modification space of biology — peptides, small molecules, antibodies, nucleic acid therapeutics — is functionally infinite. millions of viable compound × target × modification combinations. no single autonomous lab is going to cover even a meaningful fraction of it in our lifetime. @peptai_ works on disease targets through @BioProtocol infrastructure. @clarity_proto is building toward neurodegeneration with their own validation pipeline — already shipping wet-lab test batches. ALEMBIC LABS focuses on performance peptides — the MOTS-c, BPC-157, GHK-Cu, semax, retatrutide compounds biohackers actually use. these are completely different scopes within the same architectural pattern. even if two projects ended up nominally overlapping, the actual compound space is so vast that we'd be doing complementary exploration, not duplication. more open data is better open data. more agents running is better than fewer. — infrastructure deserves recognition none of these projects exist without the layers below them. and a lot of people don't see those layers clearly: @BioProtocol building the launchpad and coordination layer that makes BioDAO formation viable. @molecule_dao pioneering IP-NFT frameworks that let scientific research be funded and owned in new ways. @VitaDAO_ proving for years that decentralized longevity funding can produce real research output. @CerebrumDAO @HairDAO_ @vrials and others building specialized funds for niches that traditional grants ignore. @AdaptyvBio @GinkgoBio and other peptide synthesis and assay providers giving autonomous labs a path from in silico to wet-lab. @AnthropicAI for frontier reasoning models that make agent-based research economically viable in the first place. @BoltzAI @ChaiDiscovery for open-source structure prediction descended from AlphaFold's lineage. @biolmai for managed GPU infrastructure that lets solo builders run pharma-grade structure prediction at API prices. individually none of these are "the" breakthrough. together they're the substrate for an entirely new way of doing science. — the real competition isn't each other it's the status quo. closed datasets behind paywalls. research sitting in PDF preprints that 50 people will read. molecules biohackers use at scale with zero scientific scaffolding. multi-decade IND timelines for compounds that could be evaluated in silico in 13 minutes. every autonomous lab that ships, every dataset that gets opened, every research cycle that runs in public moves the whole ecosystem forward. that's the actual fight — not against other DeSci projects, but against the structural inertia that has kept open scientific research underfunded and slow for the last 50 years. — what i see for the next 12 months more autonomous labs. more BioDAOs forming around specific niches (gut health, hormonal optimization, cognitive enhancement, regenerative medicine, performance optimization, you name it). more wet-lab validation pipelines coming online. more grants from accelerators waking up to the model. more open datasets that compose with each other. i see a federation forming naturally. not coordinated through any single authority — just emergent, because everyone working in this space recognizes that the infrastructure benefits when everyone else's infrastructure benefits. — closing if you're building in DeSci or bio/acc — whether it's a lab, a launchpad, a fund, a synthesis partner, a community — respect to you. you're solving a hard problem in a slow-moving industry, and the only reason it's starting to feel possible is because you and others like you keep showing up. ALEMBIC LABS is one project in a much larger movement. i want to see all of us win. that's the only outcome that actually changes how science gets done. if you're building something adjacent and want to collaborate, share data, integrate, or just compare notes — DMs open. always.

Suomi
2
1
5
687
deepsy
deepsy@deepsydoin·
$alembic finally bonded. nothing changes I just keep do my job. the goal is get funds from VCs or any biotech company asap
English
3
0
7
367
dez
dez@dezpill·
@deepsydoin of course, well deserved. keep shipping we’ll help for the rest to make sure your work get eyes
English
1
0
3
42
deepsy retweetledi
WhaleWhisperer
WhaleWhisperer@Whisperer_TM·
ill take a bet on deepsy here 🧬🤖 $alembic
deepsy@deepsydoin

some thoughts on the DeSci ecosystem and where @alembiclabs fits in it. i've been thinking about this a lot since launch. the bio/acc and DeSci space is moving fast. more projects every week. real funding. real research. and a real temptation to frame everything as competition. i don't think it is. and i want to say that publicly. — the space is too big to compete in the underexplored modification space of biology — peptides, small molecules, antibodies, nucleic acid therapeutics — is functionally infinite. millions of viable compound × target × modification combinations. no single autonomous lab is going to cover even a meaningful fraction of it in our lifetime. @peptai_ works on disease targets through @BioProtocol infrastructure. @clarity_proto is building toward neurodegeneration with their own validation pipeline — already shipping wet-lab test batches. ALEMBIC LABS focuses on performance peptides — the MOTS-c, BPC-157, GHK-Cu, semax, retatrutide compounds biohackers actually use. these are completely different scopes within the same architectural pattern. even if two projects ended up nominally overlapping, the actual compound space is so vast that we'd be doing complementary exploration, not duplication. more open data is better open data. more agents running is better than fewer. — infrastructure deserves recognition none of these projects exist without the layers below them. and a lot of people don't see those layers clearly: @BioProtocol building the launchpad and coordination layer that makes BioDAO formation viable. @molecule_dao pioneering IP-NFT frameworks that let scientific research be funded and owned in new ways. @VitaDAO_ proving for years that decentralized longevity funding can produce real research output. @CerebrumDAO @HairDAO_ @vrials and others building specialized funds for niches that traditional grants ignore. @AdaptyvBio @GinkgoBio and other peptide synthesis and assay providers giving autonomous labs a path from in silico to wet-lab. @AnthropicAI for frontier reasoning models that make agent-based research economically viable in the first place. @BoltzAI @ChaiDiscovery for open-source structure prediction descended from AlphaFold's lineage. @biolmai for managed GPU infrastructure that lets solo builders run pharma-grade structure prediction at API prices. individually none of these are "the" breakthrough. together they're the substrate for an entirely new way of doing science. — the real competition isn't each other it's the status quo. closed datasets behind paywalls. research sitting in PDF preprints that 50 people will read. molecules biohackers use at scale with zero scientific scaffolding. multi-decade IND timelines for compounds that could be evaluated in silico in 13 minutes. every autonomous lab that ships, every dataset that gets opened, every research cycle that runs in public moves the whole ecosystem forward. that's the actual fight — not against other DeSci projects, but against the structural inertia that has kept open scientific research underfunded and slow for the last 50 years. — what i see for the next 12 months more autonomous labs. more BioDAOs forming around specific niches (gut health, hormonal optimization, cognitive enhancement, regenerative medicine, performance optimization, you name it). more wet-lab validation pipelines coming online. more grants from accelerators waking up to the model. more open datasets that compose with each other. i see a federation forming naturally. not coordinated through any single authority — just emergent, because everyone working in this space recognizes that the infrastructure benefits when everyone else's infrastructure benefits. — closing if you're building in DeSci or bio/acc — whether it's a lab, a launchpad, a fund, a synthesis partner, a community — respect to you. you're solving a hard problem in a slow-moving industry, and the only reason it's starting to feel possible is because you and others like you keep showing up. ALEMBIC LABS is one project in a much larger movement. i want to see all of us win. that's the only outcome that actually changes how science gets done. if you're building something adjacent and want to collaborate, share data, integrate, or just compare notes — DMs open. always.

English
0
1
5
448
deepsy
deepsy@deepsydoin·
today's work: full audit of every fold the lab has produced. pattern analysis — which research focuses produce the strongest hypotheses, where Boltz-2 confidence holds and where it doesn't, which agents pull their weight and which need rewriting. writing the findings up publicly. then shipping architecture improvements based on what the data shows. the goal isn't just more folds. it's better folds. each generation of agents informed by what the last generation surfaced. alembic gonna be my lifechanging tek, I can feel it
English
0
1
6
319